231 35 707KB
English Pages 142 Year 2009
Scuola Normale Superiore di Pisa Classe di Scienze Matematiche, Fisiche e Naturali
¨ t Zu ¨rich Universita ¨r Mathematik Institut fu
PhD Thesis The flow associated to weakly differentiable vector fields Gianluca Crippa Advisors prof. Luigi Ambrosio (Pisa)
prof. Camillo De Lellis (Z¨urich)
2005 – 2007
Contents Introduction
5
Chapter 1. The theory in the smooth framework 1. The ordinary differential equation 2. Existence and uniqueness in the classical setting 3. The classical flow of a vector field 4. The transport equation and the continuity equation
11 11 11 16 19
Part 1.
23
The Eulerian viewpoint
Chapter 2. Renormalized solutions and well-posedness of the PDE 1. A strategy to obtain uniqueness 2. Weak solutions and existence 3. Renormalized solutions 4. The DiPerna–Lions regularization scheme 5. Vector fields with Sobolev regularity 6. Vector fields with bounded variation 7. Other various renormalization results 8. Nearly incompressible vector fields
25 25 26 28 31 32 33 40 41
Chapter 3. An abstract characterization of the renormalization property 1. Forward-backward formulation 2. One-way formulation 3. Nearly incompressible vector fields
45 45 51 55
Chapter 4. Well-posedness in the two-dimensional case 1. Bounded planar divergence-free vector fields 2. Splitting on the level sets of the Hamiltonian 3. The weak Sard property 4. Structure of the level sets 5. Uniqueness on the level sets
59 59 61 63 64 65
Chapter 5. Counterexamples to the well-posedness and related constructions 1. Depauw’s counterexample 2. Oscillatory solutions to transport equations 3. Two counterexamples by DiPerna and Lions 4. Lack of propagation of regularity
69 69 70 73 78
3
4
Part 2.
CONTENTS
The Lagrangian viewpoint
79
Chapter 6. The connection between PDE and ODE 1. Pointwise uniqueness and measure valued solutions 2. The superposition principle 3. The regular Lagrangian flow 4. Ambrosio’s theory of regular Lagrangian flows 5. DiPerna–Lions’ notion of flow 6. The density of a regular Lagrangian flow 7. Bressan’s compactness conjecture
81 81 82 90 92 96 97 100
Chapter 7. A priori estimates for regular Lagrangian flows 1. A purely Lagrangian approach 2. A priori estimates for bounded vector fields and corollaries 3. Estimates for more general vector fields and corollaries 4. A direct proof of compactness 5. Lipexpp –regularity for transport equations with W 1,p coefficients 6. An application to a conjecture on mixing flows
103 103 105 114 121 125 127
Appendix: Background material and technical results 1. Measure theory 2. Lipschitz functions, extension theorems and approximate differentiability 3. Functions with bounded variation 4. Maximal functions
131 131 132 133 135
Bibliography
137
Introduction The main topic of this thesis is the study of the ordinary differential equation ( γ(t) ˙ = b(t, γ(t)) (0.1) γ : [0, T ] → Rd γ(0) = x under various regularity assumptions on the vector field b(t, x) : [0, T ] × Rd → Rd . The theory is simple and very classical in the case when b is sufficiently smooth, i.e. Lipschitz with respect to the spatial variable, uniformly with respect to the time variable. This is the so-called Cauchy–Lipschitz theory. In this case we can identify a unique flow of the vector field b, that is a map X(t, x) : [0, T ] × Rd → Rd which gathers together all trajectories, in the sense that it solves ∂X (t, x) = b(t, X(t, x)) ∂t (0.2) X(0, x) = x . Moreover, additional regularity of the vector field is inherited by the flow, as for instance regularity of x 7→ X(t, x). In this classical situation there is also a strong connection with the transport equation, an evolutionary partial differential equation which has the form ( ∂t u(t, x) + b(t, x) · ∇x u(t, x) = 0 (0.3) u : [0, T ] × Rd → R . u(0, x) = u ¯(x) When γ(t) is a solution of (0.1), the quantity g(t) = u(t, γ(t)) is constant with respect to time: indeed ∂u g(t) ˙ = (t, γ(t)) + ∇x u(t, γ(t)) · γ(t) ˙ ∂t ∂u = (t, γ(t)) + b(t, γ(t)) · ∇x u(t, γ(t)) = 0 . ∂t This means that (0.3) can be uniquely solved thanks to the theory of characteristics: the unique solution u(t, x) is the transport of the initial data u ¯(x) along solutions of (0.1), more precisely u(t, x) = u ¯ X(t, ·)−1 (x) . 5
6
Introduction
Recently a big interest has grown on the extensions of this theory to situations in which the vector field b is less regular. Apart from the theoretical importance of such an extension, the main motivation comes from the study of many nonlinear partial differential equations of the mathematical physics. In various physical models of the mechanics of fluids it is essential to deal with densities or with velocity fields which are not smooth, and this corresponds to effective real world situations: for instance the turbulent behaviour of viscous fluids or the shock waves produced by a supersonic airplane. We do not try to give here an account of the extremely wide literature on this topic, but we rather prefer to illustrate with some detail a specific case which will be presented again later on and which was a motivation for one of the main results collected in this thesis, namely Ambrosio’s theorem (see Theorem 2.12). The theory of conservation laws models situations in which the change of the amount of a physical quantity in some domain is due only to an income or an outcome of the quantity across the boundary of the domain. In the general case of several space dimensions and of a vector physical quantity these equations take the form ( ∂t u + div F (u) = 0 (0.4) u(0, ·) = u ¯, where u : [0, T ] × Rd → Rk is the unknown and F : Rk → M d×k is a given smooth map, called the flux. The well-posedness theory for these equations is presently understood only in the scalar case k = 1, and this goes back to Kruˇzkov [102], and in the one-dimensional case d = 1, via the Glimm scheme [96], or the front tracking method of Dafermos [74], or the more recent vanishing viscosity method of Bianchini and Bressan [34]. On the contrary, the general case d ≥ 2 and k ≥ 2 is presently very far from being understood. For this reason the standard approach is to tackle some particular examples, possibly in order to get some insight on the general case. For instance, consider the case in which the nonlinearity F (u) in (0.4) depends only on the modulus of the solution: d X ∂ ∂ u + f j (|u|)u = 0 t ∂xj (0.5) j=1 u(0, ·) = u ¯. This is the so-called Keyfitz and Kranzer system, introduced in [101]. For any j = 1, . . . , d the map f j : R+ → R is assumed to be smooth. A natural heuristic strategy, implemented by Ambrosio, Bouchut and De Lellis in [15] and [11], is based on a formal decoupling of (0.5) in a scalar conservation law for the modulus ρ = |u| (0.6) ∂t ρ + div f (ρ)ρ = 0 and a linear transport equation for the angular part θ = u/|u| (0.7)
∂t θ + f (ρ) · ∇θ = 0 .
Therefore, it is natural to consider weak solutions u of (0.5) such that ρ = |u| is a solution of (0.6) in the sense of entropies (the right notion of [102] which ensures
Introduction
7
existence, uniqueness and stability in the scalar case). These solutions of (0.5) are called renormalized entropy solutions and in [11] the authors showed the well-posedness of renormalized entropy solutions in the class of maps u ∈ L∞ ([0, T ] × Rd ; Rk ) such that |u| ∈ BVloc . In this case the Kruˇzkov solution ρ of (0.6) enjoys BVloc regularity, but in general it is not better, due to the creation of shocks. This means that we have to deal with the transport equation (0.7) in which the vector field f (ρ) is BVloc but not better, so that we are forced to go beyond the Cauchy–Lipschitz theory. The theory of ordinary differential equations out of the smooth context starts with the seminal work by DiPerna and Lions [83]. We want to focus now on the three main ingredients of this approach, in order to orient the reader through the material presented in this thesis; a precise description of the content of each chapter will be made later in this Introduction. In a quite different fashion with respect to the Cauchy–Lipschitz theory, the starting point is now the Eulerian problem, i.e. the well-posedness of the PDE. The first ingredient is the concept of renormalized solution of the transport equation. We say that a bounded distributional solution u of the transport equation (0.8)
∂t u + b · ∇u = 0
is a renormalized solution if for every function β ∈ C 1 (R; R) the identity (0.9) ∂t β(u) + b · ∇ β(u) = 0 holds in the sense of distributions. Notice that (0.9) holds for smooth solutions, by an immediate application of the chain-rule. However, when the vector field is not smooth, we cannot expect any regularity of the solutions, so that (0.9) is a nontrivial request when made for all bounded distributional solutions. We say that a vector field b has the renormalization property if all bounded distributional solutions of (0.8) are renormalized. The renormalization property asserts that nonlinear compositions of the solution are again solutions, or alternatively that the chain-rule holds in this weak context. The overall result which motivates this definition is that, if the renormalization property holds, then solutions of (0.3) are unique and stable. The problem is then shifted to the validity of the renormalization property. This property is not enjoyed by all vector fields: the point is that some regularity, tipically in terms of weak differentiability, is needed. The second ingredient is a series of renormalization lemmas, which ensure the renormalization property for vector fields satisfying various regularity assumptions. The typical proof ultimately relies on the regularization scheme introduced in [83] and the regularity of the vector field comes into play when one shows that the error term in the approximation goes to zero. The two most significant results of renormalization are due to DiPerna and Lions [83] and to Ambrosio [8], who deal respectively with the Sobolev and the BV case. The third ingredient is the possibility of transferring these results to the Lagrangian problem, i.e. to the ODE, establishing a kind of extended theory of characteristics. In this setting the right notion of solution of the ODE is that of regular Lagrangian flow [8], in which the “good” solution is singled out by the property that trajectories do not concentrate in small sets (this approach is slightly different from the one originally
8
Introduction
adopted in [83] to obtain results at the ODE level). The strength of this approach relies in the possibility of formulating an abstract connection between PDE and ODE. In this way, all the efforts made on the Eulerian side are automatically translated into well-posedness results for the regular Lagrangian flow. We now illustrate the contribution of the author to the development of this research area. The following papers, with the exception of the first one, have been produced during the author’s PhD studies in 2005–2007 at the Scuola Normale Superiore di Pisa, with a co-supervision at the Universit¨at Z¨ urich. In a joint work with Luigi Ambrosio and Stefania Maniglia [14] we study some fine properties of normal traces of vector fields with measure divergence and we apply them to show the renormalization property for special vector fields with bounded deformation. Moreover, we prove approximate continuity properties of solutions of transport equations. This work is not presented in detail in this thesis, since it has already been presented in the Tesi di Laurea [64]. We mention it in Section 2.7 for the results relative to the renormalization property and in Section 6.7 for some more recent progresses based on the approximate continuity results. The paper with Camillo De Lellis [65] settles, in the negative, a conjecture made by Bressan about the possibility of applying transport equation techniques to the study of multi-dimensional systems of conservation laws, in the same spirit of the derivation mentioned for the Keyfitz and Kranzer system. This is briefly presented in Section 5.2. In the two-dimensional divergence-free case the presence of a Hamiltonian function which is (at least formally) conserved along the flow gives an additional structure to the problem and this allows a dramatic reduction of the regularity needed for the uniqueness. In a first paper with Ferruccio Colombini and Jeffrey Rauch [57] we show that all the results contained in the literature on this problem hold also in the case of bounded, and not necessarily zero, divergence. However, a technical assumption of unclear meaning is present in all these results. A further investigation, in a work in progress in collaboration with Giovanni Alberti and Stefano Bianchini [5], identifies the sharpest hypothesis needed for the uniqueness and clarifies its meaning. A new technique of dimensional reduction via a splitting on the level sets of the Hamiltonian is introduced. The positive result, which is presented in a basic case in Chapter 4, is complemented by two counterexamples which show the necessity of the “extra assumption”. In a joint work with Fran¸cois Bouchut [37] the meaning of the renormalization property is discussed: its connections with forward and backward uniqueness, strong continuity and smooth approximations of the solution are analysed. Chapter 3 is completely devoted to the presentation of these results. A second paper in collaboration with Camillo De Lellis [66] (see also the conference proceedings [67]) introduces a new approach to the theory of regular Lagrangian flows, in the W 1,p case with p > 1. This approach is based on some a priori quantitative estimates, which can be derived directly at the ODE level, with no mention to the PDE theory. These estimates allow to recover in a direct way various known wellposedness results, but also have some new interesting corollaries, in terms of regularity and compactness of the regular Lagrangian flow. This work is presented in Chapter 7.
Introduction
9
Eventually we mention the lecture notes [13] in collaboration with Luigi Ambrosio: they present a self-contained survey of the theory of regular Lagrangian flows, with a particular interest to the recent developments.
We finally illustrate the content of the various chapters of this thesis. For more details the reader is referred to the introduction of each chapter. In Chapter 1 we present the Cauchy–Lipschitz theory: we start with the classical Picard–Lindel¨of theorem regarding existence and uniqueness for ODEs with Lipschitz vector field and we present some variants and extensions. Then we prove existence, uniqueness and regularity of the classical flow of a vector field. Finally we present some ideas of the theory of characteristics and the connection between the ODE and the transport and the continuity equations in the smooth case. Starting from Chapter 2 we begin to consider the non-smooth case from the Eulerian viewpoint. We first introduce heuristically the motivation for the concept of renormalized solution. Then we define rigorously the notion of weak (distributional) solution and we see that existence holds under very general hypotheses. The renormalization property is then introduced, together with its consequences for the well-posedness of the PDE. We describe the DiPerna–Lions regularization scheme and the proof of the renormalization property by DiPerna and Lions in the Sobolev case and by Ambrosio in the BV case. Finally we comment on other various renormalization results and on the case of nearly incompressible vector fields. Chapter 3 is devoted to [37]: we show, first in the divergence-free framework and then in more general cases, the relations between the notions of renormalization, strong continuity of the solution with respect to the time, forward and backward uniqueness in the Cauchy problem and approximation (in the sense of the norm of the graph of the transport operator) of the solution with smooth maps. The two-dimensional case is analysed in detail in Chapter 4, in which we mainly follow [5]. In the basic case of an autonomous bounded divergence-free vector field we show that, assuming a weak Sard property on the Hamiltonian, the two-dimensional PDE can be split on the level sets of the Hamiltonian. Then we show uniqueness “line by line”, using explicit arguments and a careful study of the structure of the level sets, and this eventually implies uniqueness for the full two-dimensional equation. Chapter 5 collects many counterexamples to the uniqueness, in particular the examples of [81] and [83] which shed light on the necessity of the assumptions made. Moreover, some related constructions are presented, in particular the counterexample of [65], which shows the impossibility of building a theory for multi-dimensional systems of conservation laws based on transport equations, and the examples of [62], relative to the lack of propagation of regularity in the transport equation. With Chapter 6 we start to present the Lagrangian viewpoint. First, the equivalence between pointwise uniqueness for the ODE and uniqueness of positive measure-valued solutions of the continuity equation is shown, exploiting the superposition principle for solutions of the PDE. Then, we introduce the concept of regular Lagrangian flow and we show how the uniqueness of bounded solutions of the continuity equation implies existence and uniqueness of the regular Lagrangian flow. We finally comment on the
10
Introduction
DiPerna–Lions’ notion of flow and on the case of nearly incompressible vector fields, with particular attention to a compactness conjecture made by Bressan. Chapter 7 is devoted to the a priori quantitative estimates for Sobolev vector fields [66]. The estimates are derived first in the bounded case, and then assuming more general growth conditions. The consequences regarding existence, uniqueness, stability (in a quantitative form), compactness and regularity of the regular Lagrangian flow are presented. Some applications to the propagation of a mild regularity in the transport equation and to a mixing conjecture made by Bressan are also addressed. Finally, in the Appendix we recall some basic tools and results which are widely used throughout the whole thesis. As a general warning to the reader we observe that in many cases we do not state our results under the sharpest possible assumptions: this is done in order to simplify the exposition and to avoid annoying technicalities. For instance, in almost all the thesis we deal only with globally bounded vector fields, but more general growth conditions would be sufficient, compare Section 7.3. All the statements are typically completely rigorous, while concerning the proofs we have sometimes chosen to skip some details, to merely specify the main steps or to enlighten some important or delicate points, at the expense of a more complete proof. This is done according to the author’s taste and not to the absolute importance of each result. For the main notation used we refer the reader to the Appendix or to the first place where each symbol or locution appears.
CHAPTER 1
The theory in the smooth framework In this chapter we illustrate the main results of the theory of ordinary differential equations in the smooth framework. Most of these results are very classical, hence we give only a sketch of the proofs. In Sections 1.1 and 1.2 we study existence, uniqueness and stability properties of the solutions. In Section 1.3 we introduce the concept of flow of a vector field. In Section 1.4 we illustrate the link between the ordinary differential equation and two partial differential equations, namely the transport equation and the continuity equation. Basic references for these topics are for instance [18], [39], [93], [97] and [120]. 1. The ordinary differential equation Let b : D ⊂ Rt × Rdx → Rd be a continuous time-dependent vector field, where D is an open set. A (classical) solution of the ordinary differential equation (1.1)
γ(t) ˙ = b(t, γ(t))
consists of an interval [t1 , t2 ] ⊂ R and a function γ ∈ C 1 ([t1 , t2 ]; Rd ) which satisfies (1.1) for every t ∈ [t1 , t2 ]. In particular this implies that (t, γ(t)) ∈ D for every t ∈ [t1 , t2 ]. We also say that γ is an integral curve or a characteristic curve of the vector field b. Notice that, if b ∈ C ∞ (D; Rd ), then every characteristic curve of b is in C ∞ ([t1 , t2 ]; Rd ). Now we fix (t0 , x0 ) ∈ D and we consider the Cauchy problem for the ordinary differential equation: ( γ(t) ˙ = b(t, γ(t)) (1.2) γ(t0 ) = x0 . A solution to this problem is a function γ ∈ C 1 ([t1 , t2 ]; Rd ) which is a characteristic curve of b and satisfies the condition γ(t0 ) = x0 . In particular we must require t0 ∈ [t1 , t2 ]. It is immediate to check that γ ∈ C 1 ([t0 − r, t0 + r]; Rd ) is a solution of (1.2) if and only if γ ∈ C 0 ([t0 − r, t0 + r]; Rd ) and satisfies the identity Z t γ(t) = x0 + b(s, γ(s)) ds for every t ∈ [t0 − r, t0 + r]. t0
2. Existence and uniqueness in the classical setting We want to discuss the well-posedness of (1.2) under the assumption that the vector field b is sufficiently smooth. The following simple example shows that the situation could be complicated even in the case of quite simple one-dimensional vector fields. 11
12
1. THE THEORY IN THE SMOOTH FRAMEWORK
Example 1.1 (The square rootp example). We consider on R the (continuous but not Lipschitz) vector field b(t, x) = |x|. The ordinary differential equation ( p γ(t) ˙ = |γ(t)| γ(0) = 0 has the solution
c
γ (t) =
1 4 (t
0 if t ≤ c − c)2 if t ≥ c
for every value of the parameter c ∈ [0, +∞]. The solution can “stay at rest” in the origin for an arbitrary time. This example will be considered again later on: see Remark 6.6. This example stresses that some smoothness assumptions are necessary to obtain uniqueness. The following classical result by Picard–Lindel¨of requires Lipschitz continuity of b with respect to the spatial variable, with some uniformity with respect to the time. Theorem 1.2 (Picard–Lindel¨of). Let b be a continuous vector field defined on an open set containing the rectangle n o D = (t, x) ∈ R × Rd : |t − t0 | ≤ α , |x − x0 | ≤ β . Assume that b is Lipschitz continuous with respect to the spatial variable, uniformly with respect to the time, in D, i.e. (1.3)
|b(t, x) − b(t, y)| ≤ K|x − y|
for every (t, x), (t, y) ∈ D,
and let M be such that |b(t, x)| ≤ M in D. Then there exists a unique solution γ ∈ C 1 ([t0 − r, t0 + r]; Rd ) to (1.2), where β 1 r < min α, , . M K Proof. Choosing r as in the statement we define Ir = [t0 − r, t0 + r]. We will show that there exists a unique γ ∈ C 0 (Ir ; Rd ) such that Z t γ(t) = x0 + b(s, γ(s)) ds for every t ∈ Ir . t0
We want to identify a complete metric space X on which the operator Z t T [γ](t) = x0 + b(s, γ(s)) ds t0
is a contraction. We set X = γ ∈ C 0 (Ir ; Rd ) : γ(t0 ) = x0 and |γ(t) − x0 | ≤ β for every t ∈ Ir . X is a closed subset of C 0 (Ir ; Rd ) equipped with the norm of the uniform convergence, hence it is a complete metric space. If γ ∈ X then clearly T [γ] is a continuous map
2. EXISTENCE AND UNIQUENESS IN THE CLASSICAL SETTING
13
which satisfies T [γ](t0 ) = x0 . Moreover Z t |T [γ](t) − x0 | ≤ |b(s, γ(s))| ds ≤ M |t − t0 | ≤ M r < β . t0
This implies that T takes value in X. We now check that T is a contraction. Take γ1 and γ2 in X. Then Z t |T [γ1 ](t) − T [γ2 ](t)| ≤ |b(s, γ1 (s)) − b(s, γ2 (s))| ds t0 Z t ≤ K |γ1 (s) − γ2 (s)| ds t0
≤ K|t − t0 |kγ1 − γ2 kL∞ (Ir ) ≤ Krkγ1 − γ2 kL∞ (Ir ) . This implies that kT [γ1 ] − T [γ2 ]kL∞ (Ir ) ≤ Krkγ1 − γ2 kL∞ (Ir ) and since Kr < 1 we can apply the Banach fixed point theorem to get the existence of a unique fixed point for T . This is precisely the unique solution to (1.2). Remark 1.3. This result is also valid if we assume that b is only summable with respect to the time variable and Lipschitz continuous in x uniformly in t. For a proof see for instance Theorem 4.0.7 of [22]. If we drop the Lipschitz continuity assumption, then we lose uniqueness of the solution: compare with Example 1.1. However, we still have existence, thanks to the following result. Theorem 1.4 (Peano). Let b be a continuous and bounded vector field defined on an open set containing the rectangle n o D = (t, x) ∈ R × Rd : |t − t0 | ≤ α , |x − x0 | ≤ β . Then there exists a local solution to (1.2). Proof. We take r < min{α, β/M } and consider again the operator T and the space X defined in the proof of Theorem 1.2. X is a non-empty bounded convex closed subset of C 0 (Ir ; Rd ) equipped with the norm of the uniform convergence and T : X → X is a linear continuous operator (because of the uniform continuity of b). Moreover for every γ ∈ X we can estimate Z t 0 |T [γ](t) − T [γ](t )| = |b(s, γ(s))| ds ≤ M |t − t0 | . 0 t
Hence by the Ascoli–Arzel`a theorem T is a compact operator. Then the existence of a fixed point in X is ensured by the Caccioppoli–Schauder fixed point theorem (see Corollary 11.2 of [94]). Remark 1.5. In the one-dimensional autonomous case Rwe have (local) uniqueness x dy provided b is continuous and b(x0 ) 6= 0. Indeed, set G(x) = x0 b(y) . Since b(x0 ) 6= 0 we
14
1. THE THEORY IN THE SMOOTH FRAMEWORK
can find a neighbourhood of x0 in which G is of class C 1 with G0 6= 0; hence G has a C 1 inverse. If γ is a solution of (1.2) defined in a neighbourhood of 0 we can compute γ(t) ˙ d G(γ(t)) = = 1, dt b(γ(t)) which eventually implies G(γ(t)) = t and thus γ(t) = G−1 (t). Since G only depends on b, this means that the solution is unique. This is only true in a neighbourhood of t = t0 : similarly to the case described in Example 1.1, the solution of the problem ( p γ(t) ˙ = |γ(t)| γ(0) = x0 < 0 reaches the origin (and up to that time it is unique), but it can stay at rest there for an arbitrary amount of time before leaving it. We now present two other conditions which are a bit more general than the Lipschitz condition required in Theorem 1.2, but nevertheless they are sufficient to get uniqueness. Proposition 1.6 (One-sided Lipschitz condition). Uniqueness forward in time for (1.2) holds if the Lipschitz continuity condition (1.3) in Theorem 1.2 is replaced by the following one-sided Lipschitz condition: b(t, x) − b(t, y) · (x − y) ≤ K|x − y|2 for every (t, x), (t, y) ∈ D. By “forward in time” we mean the following: two solutions γ1 and γ2 to (1.2), which by definition satisfy γ1 (t0 ) = γ2 (t0 ) = x0 , coincide for t > t0 . Observe that the one-sided Lipschitz condition cannot guarantee uniqueness backward in time, since it is not invariant under an inversion of the sign of t in the equation, which amounts to a change of sign of b. Proof of Proposition 1.6. Consider r(t) = |γ1 (t) − γ2 (t)|2 and notice that r(t0 ) = 0. Using the one-sided Lipschitz condition we can estimate r(t) ˙ = 2 γ1 (t) − γ2 (t) · b(t, γ1 (t)) − b(t, γ2 (t)) ≤ K|γ1 (t) − γ2 (t)|2 = Kr(t) . This implies that i dh r(t)e−Kt ≤ 0 , dt hence r(t)e−Kt ≤ r(t0 )e−Kt0 = 0 that is the desired thesis.
for every t ≥ t0 ,
Proposition 1.7 (Osgood condition). Uniqueness for (1.2) holds if the Lipschitz continuity condition (1.3) in Theorem 1.2 is replaced by the following Osgood condition: |b(t, x) − b(t, y)| ≤ ω(|x − y|)
for every (t, x), (t, y) ∈ D,
where ω : R+ → R+ is an increasing function satisfying ω(0) = 0, ω(z) > 0 for every z > 0 and Z 1 1 dz = +∞ . ω(z) 0
2. EXISTENCE AND UNIQUENESS IN THE CLASSICAL SETTING
15
Proof. As in the proof on the previous proposition we define r(t) = |γ1 (t)−γ2 (t)|2 . We have r(t0 ) = 0 and we assume by contradiction that r(t) > 0 for t ∈]t0 , t0 + r[ for some r > 0. This local forward uniqueness clearly implies global forward uniqueness, and backward global uniqueness is proved in the same way. We compute r(t) ˙ = 2 γ1 (t) − γ2 (t) · γ˙ 1 (t) − γ˙ 2 (t) = 2 γ1 (t) − γ2 (t) · b(t, γ1 (t)) − b(t, γ2 (t)) . Using the Osgood condition and integrating with respect to the time we get Z tp r(s) ω(r(s)) ds . (1.4) r(t) ≤ 2 t0
p
We notice that r(s) = |γ1 (t) − γ2 (t)| ∈ L1 ([t0 , t0 + r]) and we apply an extension of the Gronwall lemma, which goes as follows. Define Z tp R(t) = 2 r(s) ω(r(s)) ds . t0
Since we are assuming that r(t) > 0 for t ∈]t0 , t0 + r[ we deduce that R(t) > 0 for t ∈]t0 , t0 + r[. Moreover applying (1.4) we have p p ˙ R(t) = r(t) ω(r(t)) ≤ r(t) ω(R(t)) , from which we get (1.5)
˙ p R(t) ≤ r(t) ω(R(t))
t > t0 .
We set Z
1
1 dr z ω(r) and we observe that Ω(z) is well-defined for every z > 0, satisfies Ω0 (z) = −1/ω(z) and Ω(z) ↑ +∞ as z ↓ 0, because of the Osgood condition. Integrating (1.5) we deduce that for every t0 < s < t we have Z tp r(τ ) dτ , −Ω(R(t)) + Ω(R(s)) ≤ Ω(z) =
s
which however cannot hold for s very close to t0 . Indeed, since R(s) ↓ 0 as s ↓ t0 , in this case we would have Ω(R(s)) ↑ +∞ as s ↓ t0 , while the integral on the right hand side stays finite. We conclude this section with a discussion about the maximal interval of existence of the solution to (1.2). The solution that we constructed in the previous theorems is in fact local in time. However, if b : D ⊂ R × Rd → Rd is continuous and bounded, then every solution γ :]t1 , t2 [→ Rd of (1.2) can be extended to the closed interval [t1 , t2 ]. Indeed, for every t1 < t < t0 < t2 we have Z t0 0 |γ(t ) − γ(t)| ≤ |b(s, γ(s))| ds ≤ M |t0 − t| , t
16
1. THE THEORY IN THE SMOOTH FRAMEWORK
where M is an upper bound for |b| on D. Hence γ is Lipschitz and admits a unique extension to the closure of its domain of definition. Now notice that, if (t1 , γ(t1 )) or (t2 , γ(t2 )) is not on the boundary of D, we can apply again the local existence result proved in Theorem 1.2 (or in Theorem 1.4). We conclude the following result: Theorem 1.8. Let b : D ⊂ R × Rd → Rd be a continuous and bounded vector field. Assume that b is Lipschitz continuous with respect to the spatial variable, uniformly with respect to the time, on every bounded rectangle contained in D. Fix (t0 , x0 ) ∈ D. Then there exists a unique solution to (1.2) and it can be extended until its graph touches the boundary of D. This identifies a maximal interval of existence for the solution to (1.2). If b is globally defined and bounded we deduce Corollary 1.9. Let b : I × Rd → Rd be a continuous and bounded vector field, where I ⊂ R is an interval. Assume that b is locally Lipschitz continuous with respect to the spatial variable, uniformly with respect to the time. Then, for every (t0 , x0 ) ∈ I ×Rd there exists a unique solution to (1.2) which is defined for t ∈ I. 3. The classical flow of a vector field We want to compare two solutions γ and γ b to the ordinary differential equation, with initial data at time t0 respectively equal to x0 and x c0 . We can compute Z t Z t γ(t) − γ b(t) = x0 − x c0 + b(s, γ(s)) ds − b(s, γ b(s)) ds t0 t
Z = x0 − x c0 +
t0
t0
b(s, γ(s)) − b(s, γ b(s)) ds .
Hence we deduce Z |γ(t) − γ b(t)| ≤ |x0 − x c0 | + Lip(b)
t
t0
|γ(s) − γ b(s)| ds ,
where Lip(b) denotes, as usual, the Lipschitz constant of b. A simple Gronwall argument then gives (1.6) |γ(t) − γ b(t)| ≤ |x0 − x c0 | exp K|t − t0 | . This means that the solution depends Lipschitz continuously on the initial data. A very similar argument shows that, if we consider two solutions γ and γ b with the b same initial data x0 but relative to different vector fields b and b, we have the continuity estimate (1.7) |γ(t) − γ b(t)| ≤ |t − t0 |kb − bbk∞ exp K|t − t0 | . Looking at the solution of (1.2) as a function not only of the time but also of the initial point x0 we are led to the following definition:
3. THE CLASSICAL FLOW OF A VECTOR FIELD
17
Definition 1.10 (Classical flow of a vector field). Consider a continuous and bounded vector field b : I × Rd → Rd , where I ⊂ R is an interval, and let t0 ∈ I. The (classical) flow of the vector field b starting at time t0 is a map X(t, x) : I × Rd → Rd which satisfies (1.8)
∂X (t, x) = b(t, X(t, x)) ∂t X(t0 , x) = x .
If the vector field b is bounded and Lipschitz with respect to the space variable we immediately deduce existence and uniqueness of the flow from Corollary 1.9. Moreover (1.6) also gives Lipschitz regularity of the flow. Corollary 1.11. Let b : I × Rd → Rd be a continuous and bounded vector field, where I ⊂ R is an interval. Assume that b is locally Lipschitz continuous with respect to the spatial variable, uniformly with respect to the time. Then, for every t0 ∈ I there exists a unique classical flow of b starting at time t0 . Moreover, the flow is Lipschitz continuous with respect to t and x. Now we want to show that additional smoothness of b implies that the flow is also smooth with respect to the spatial variable. Assume that b is C 1 with respect to the spatial variable, uniformly with respect to the time. We first discuss the differentiability in the direction given by a unit vector e ∈ Sd−1 . For every small h ∈ R we need to compare X(t, x) and X(t, x + he). We first observe that differentiating formally (1.8) with respect to x in the direction e we obtain the following ordinary differential equation for Dx X(t, x)e: ∂ Dx X(t, x)e = Dx b (t, X(t, x))Dx X(t, x)e . ∂t Motivated by this, we define we (t, x) to be the solution of ∂we (t, x) = Dx b (t, X(t, x))we (t, x) (1.9) ∂t w (t , x) = e . e 0 This is a linear ordinary differential equation which depends on the parameter x ∈ Rd . It is readily checked that for every x ∈ Rd there exists a unique solution we (t, x) defined for t ∈ I. Moreover it is simple to prove that we (t, x) depends continuously on the parameter x ∈ Rd . We claim that X(t, x + he) − X(t, x) (1.10) → we (t, x) as h → 0. h This will give Dx X(t, x)e = we (t, x) and, since we (t, x) is continuous in x, we will get that the flow X(t, x) is differentiable with respect to x with continuous differential. Going back to (1.9) we also deduce that, if b is C k with respect to the spatial variable, then the flow X is C k with respect to x.
18
1. THE THEORY IN THE SMOOTH FRAMEWORK
We now prove (1.10). We define the map X(t, x + he) − X(t, x) . h We notice that ze,h (0, x) = e and we compute i ∂ze,h 1 ∂X ∂X 1h (t, x) = (t, x + he) − (t, x) = b(t, X(t, x + he)) − b(t, X(t, x)) ∂t h ∂t ∂t h Z 1 = Dx b t, τ X(t, x + he) + (1 − τ )X(t, x) dτ ze,h (t, x) 0 = Dx b t, X(t, x) ze,h (t, x) + Ψe,h (t, x)ze,h (t, x) , ze,h (t, x) =
where we have set Z 1h i Ψe,h (t, x) = Dx b t, τ X(t, x + he) + (1 − τ )X(t, x) − Dx b t, X(t, x) dτ . 0
Since we are assuming that the vector field b is C 1 with respect to the spatial variable we deduce that Ψe,h (t, x) → 0
as h → 0, uniformly in t and x.
We consider ϕe,h (t, x) = ze,h (t, x) − we (t, x) and we notice that ϕe,h satisfies ∂ϕe,h (t, x) = Dx b t, X(t, x) ϕe,h (t, x) + Ψe,h (t, x)ze,h (t, x) ∂t ϕ (0, x) = 0 . e,h
Recalling (1.7) and the fact that |Ψe,h (t, x)| = o(1) and |ze,h (t, x)| = O(1) we deduce that |ϕe,h (t, x)| = o(1). Going back to the definitions of ϕe,h and of ze,h we see that this is precisely (1.10). We conclude the following theorem, improving the result of Corollary 1.11. Theorem 1.12. Let b : I × Rd → Rd be a smooth and bounded vector field, where I ⊂ R is an interval. Then for every t0 ∈ I there exists a unique classical flow of b starting at time t0 , which is smooth with respect to t and x. Notice also that, for every t ∈ I, the map X(t, ·) : Rd → Rd is a smooth diffeomorphism. The Jacobian J(t, x) = det(∇x X(t, x)) satisfies the equation ∂J (t, x) = (div b)(t, X(t, x))J(t, x) , (1.11) ∂t which also implies that J(t, x) > 0 for every t ∈ I. Moreover, using the notation X(t, s, x) for the flow of b starting at time s ∈ I, the following semi-group property holds, as a consequence of the uniqueness of the flow: (1.12)
X(t2 , t0 , x) = X(t2 , t1 , X(t1 , t0 , x))
for every t0 , t1 , t2 ∈ I.
4. THE TRANSPORT EQUATION AND THE CONTINUITY EQUATION
19
4. The transport equation and the continuity equation In this section we show the relation between the ordinary differential equation and the transport equation (1.13)
∂t u(t, x) + b(t, x) · ∇u(t, x) = 0 ,
u(t, x) : [0, T ] × Rd → R ,
and the continuity equation ∂t µ + div bµ = 0 ,
(1.14)
where µ = {µt }t∈[0,T ] is a family of locally finite signed measures on Rd , which depends on the time parameter t ∈ [0, T ]. The continuity equation is intended in distributional sense, according to the following definition. Definition 1.13. A family µ = {µt }t∈[0,T ] of locally finite signed measures on Rd is a solution of the continuity equation (1.14) if Z TZ ∂t ϕ(t, x) + b(t, x) · ∇ϕ(t, x) dµt (x) dt = 0 0
Rd
for every test function ϕ ∈ Cc∞ (]0, T [×Rd ). We recall here only the main ideas of the theory of characteristics for partial differential equations and we derive some explicit formulas for the solution in the smooth case. The connection between the Lagrangian problem (ODE) and the Eulerian problem (PDE) will be investigated in detail in Chapter 6, in which a kind of extension of the theory of characteristics to the non-smooth case will be described. If we consider a characteristic curve γ(t) of the vector field b(t, x) and a smooth solution u(t, x) of (1.13) we can easily check that the quantity g(t) = u(t, γ(t)) is constant with respect to time. Indeed ∂u (t, γ(t)) + ∇x u(t, γ(t)) · γ(t) ˙ ∂t ∂u = (t, γ(t)) + b(t, γ(t)) · ∇x u(t, γ(t)) = 0 . ∂t
g(t) ˙ =
This means that the solution u is constant on the characteristic lines of b. Hence, if we couple the transport equation (1.13) with an initial data u(0, ·) = u ¯, we expect u(t, x) = u ¯ X(t, ·)−1 (x) = u ¯ X(0, t, x) to be a solution of the Cauchy problem. This is easily checked with a direct computation, observing that the flow X(t, s, x) satisfies the equation (1.15)
∂X (t, s, x) + b(s, x) · ∇x X(t, s, x) = 0 . ∂s
This is also the unique solution with initial data u ¯, since we already know that every solution has to be constant along characteristics. We then conclude the following proposition.
20
1. THE THEORY IN THE SMOOTH FRAMEWORK
Proposition 1.14. If the vector field b and the initial data u ¯ are C 1 , then the transport equation (1.13) has the unique solution u(t, x) = u ¯ X(t, ·)−1 (x) . Similarly, in the case of a transport equation with a source term f ∈ C 1 on the right hand side ∂t u(t, x) + b(t, x) · ∇u(t, x) = f (t, x) , we have the explicit expression Z (1.16)
t
f (s, X(s, t, x)) ds .
u(t, x) = u ¯(X(0, t, x)) + 0
We have a similar result for the continuity equation (1.14), coupled with the initial data µ0 = µ ¯ ∈ M(Rd ). Proposition 1.15. Assume that the vector field b is C 1 . Then, for any initial data µ ¯, the solution of the continuity equation (1.14) is given by Z Z (1.17) µt = X(t, ·)# µ ¯, i.e. ϕ dµt = ϕ(X(t, x)) d¯ µ(x) ∀ϕ ∈ Cc (Rd ), Rd
Rd
where X(t, ·)# µ ¯ denotes the push-forward of the measure µ ¯ via the map X(t, ·) : Rd → d R , defined according to (A.1). Proof. Notice first that we need only to check the distributional identity ∂t µ + div (bµ) = 0 on test functions of the form ψ(t)ϕ(x), that is Z Z Z 0 (1.18) ψ (t)hµt , ϕi dt + ψ(t) b(t, x) · ∇ϕ(x) dµt (x) dt = 0 . R
Rd
R
We notice that the map Z t 7→ hµt , ϕi =
ϕ(X(t, x)) d¯ µ(x) Rd
belongs to C 1 ([0, T ]), since the flow X is C 1 with respect to the time variable. In order R to check (1.18) we need to show that the distributional derivative of this map is Rd b(t, x) · ∇ϕ(x) dµt (x), but by the C 1 regularity we only need to compute the pointwise derivative. Since the flow satisfies ∂X (t, x) = b(t, X(t, x)) ∂t for every t and x we can deduce Z d d hµt , ϕi = ϕ(X(t, x)) d¯ µ(x) dt dt d Z R = ∇ϕ(X(t, x)) · b(t, X(t, x)) d¯ µ(x) Rd
= hb(t, ·)µt , ∇ϕi , hence we have shown the desired thesis.
4. THE TRANSPORT EQUATION AND THE CONTINUITY EQUATION
21
When the initial data is a locally summable function, i.e. µ ¯ = ρ¯L d , we can give an explicit expression for the density ρt of µt : ρ¯ ρt (x) = ◦ X(t, ·)−1 (x) . det(∇X(t, ·)) This is a consequence of (1.17) and of the area formula; for a derivation of this equality see Section 3 of [13].
Part 1
The Eulerian viewpoint
CHAPTER 2
Renormalized solutions and well-posedness of the PDE In this chapter we begin to investigate the well-posedness of the transport equation out of the smooth setting. In the first section we present some formal computations in order to motivate the definition of renormalized solution, which will be central in the theory. In Section 2.2 we introduce the weak formulation of the transport equation, which is needed in order to consider the non-smooth case; it is also shown a very general result of existence of bounded solutions. In Sections 2.3 and 2.4 we introduce the notion of renormalized solution and we enlighten its importance in the well-posedness theory; we also propose a general strategy, due to DiPerna and Lions [83], which is one of the main roads to the proof of the renormalization property and thus of the well-posedness. In Sections 2.5 and 2.6 we illustrate the proof of the renormalization property in two important situations, namely the Sobolev case (DiPerna and Lions, [83]) and the BV case (Ambrosio, [8]). In the last two sections we finally present some other renormalization results due to various authors and the case of nearly incompressible vector fields.
1. A strategy to obtain uniqueness In this section we describe informally one of the main tools that can be used to show the well-posedness of the transport equation ( ∂t u(t, x) + b(t, x) · ∇u(t, x) = 0 (2.1) u(0, x) = u ¯(x) . In order to illustrate the motivation for the concept of renormalized solution we present some computations. We proceed in a merely formal way: the argument cannot be justified without regularity assumptions on the vector field b, the uniqueness result being false in this general context (see the counterexample in Section 5.1 and the more recent ones of [5]). However, this presentation has the advantage of making evident the main points of the subsequent analysis. We start from (2.1) and we multiply it by 2u, obtaining 2u∂t u + 2ub · ∇u = 0 . Then we rewrite this equality as (2.2)
∂t u2 + b · ∇u2 = 0 . 25
26
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
Now we assume that div b = 0 and integrate over all Rd , for every t ∈ [0, T ] fixed. This gives Z Z d 2 div b(t, x)u(t, x)2 dx = 0 , u(t, x) dx = − dt Rd Rd applying the divergence theorem (we also assume a sufficiently fast decay at infinity). This implies that the L2 (Rd ) norm is conserved: d ku(t, ·)kL2 (Rd ) = 0 . dt But the transport equation (2.1) is linear: then, in order to show uniqueness, it is enough to show that if the initial data is u ¯ ≡ 0, then the only solution is u ≡ 0. But this is clearly implied by the conservation of the L2 (Rd ) norm. This formal argument has in fact two gaps. First, the application of the chain-rule to obtain (2.2) is not justified: indeed, solutions of (2.1) are not smooth in general, hence we cannot simply write (2.3)
∂t u2 = 2u∂t u
and
∇u2 = 2u∇u .
The second gap is a bit more hidden, but it is also extremely relevant. The initial data in (2.1) is meant in the sense of distributions (see (2.4)). However, in order to apply (2.3) to show that ku(t, ·)kL2 (Rd ) = 0 for every t, we need to know a priori that ku(t, ·)kL2 (Rd ) → k¯ ukL2 (Rd )
as t ↓ 0,
but this is a property of strong continuity of the solution that cannot be deduced only from the weak formulation: see again the example presented in Section 5.1. 2. Weak solutions and existence We first introduce the weak formulation of the transport equation (2.1). Let b : [0, T ] × Rd → Rd be a vector field and denote by div b the divergence of b (with respect to the spatial coordinates) in the sense of distributions. Definition 2.1. If b and u ¯ are locally summable functions such that the distributional divergence of b is locally summable, then we say that a locally bounded function u : [0, T ] × Rd → R is a (weak) solution of (2.1) if the following identity holds for every function ϕ ∈ Cc∞ ([0, T [×Rd ): Z TZ Z h i (2.4) u ∂t ϕ + ϕdiv b + b · ∇ϕ dxdt = − u ¯(x)ϕ(0, x) dx . 0
Rd
Rd
Notice that this is the standard notion of weak solution of a PDE. For smooth solutions, equation (2.4) can be immediately deduced from (2.1) multiplying it by ϕ and integrating by parts. d We can equivalently define weak solutions noticing that, if u ∈ L∞ loc ([0, T ]×R ), then 1 ∂t u has a meaning as a distribution. Moreover, assuming that div b ∈ Lloc ([0, T ] × Rd ), we can define the product b · ∇u as a distribution via the equality hb · ∇u, ϕi = −hbu, ∇ϕi − hudiv b, ϕi
for every ϕ ∈ Cc∞ (]0, T [×Rd ).
2. WEAK SOLUTIONS AND EXISTENCE
27
This enables us to give directly a distributional meaning to the transport equation (2.1). The initial data of the Cauchy problem (2.1) can be recovered using the following observation. Remark 2.2 (Weak continuity in time). If we test the transport equation (2.1) against test functions of the form ψ(t)ϕ(x) with ψ ∈ Cc∞ (]0, T [) and ϕ ∈ Cc∞ (Rd ) we easily obtain that Z Z d u(t, x)ϕ(x) dx = u(t, x) ϕ(x)div b(t, x) + b(t, x) · ∇ϕ(x) dx dt Rd Rd in the sense of distributions over [0, T ]. This implies in particular the estimate Z d (2.5) dt d u(t, x)ϕ(x) dx ≤ kϕkC 1 (Rd ) V (t) , R
where
Z
|div b(t, x)| + |b(t, x)| dx ∈ L1 ([0, T ]) . Rd R Let Lϕ ⊂ [0, T ] be the set of the Lebesgue points of the map t 7→ u(t, x)ϕ(x) dx. We know that L 1 ([0, T ] \ Lϕ ) = 0. Now consider a countable set Z which is dense in Cc1 (Rd ) with respect to the C 1 norm and set LZ = ∩ϕ∈Z Lϕ . Clearly we have L 1 ([0, T ] \ LZ ) = 0. The restriction of u(t, ·) to LZ provides a uniformly continuous family of bounded functionals on Cc1 (Rd ), since estimate (2.5) implies Z Z Z t u(t, x)ϕ(x) dx − u(s, x)ϕ(x) dx ≤ kϕkC 1 V (τ ) dτ . V (t) = kuk∞
Rd
Rd
s
Therefore to a continuous curve u ˜(t, ·) in 1 d 0 u(t, ·) can be extended in a unique way ∞ d Cc (R ) , for t ∈ [0, T ]. But since u(t, x) ∈ L ([0, T ] × R ) this also implies that 0 u ˜(t, ·) is a continuous curve in L1 (Rd ) . From the above argument, which is indeed classical in the theory of evolutionary PDEs (see for instance Lemma 8.1.2 of [18] or Theorem 4.1.1 of [74]), we deduce that, up to a modification of u(t, ·) in a negligible set of times, we can assume that t 7→ u(t, ·) is weakly∗ continuous from [0, T ] into L∞ (Rd ). A similar remark applies to the continuity equation (1.14): we can assume that the map t 7→ µt is weakly∗ continuous from [0, T ] to M(Rd ). It follows in particular that u(t, ·) and µt are defined for every t ∈ [0, T ], and in particular at the endpoints; this also gives a sense to the Cauchy data at t = 0. As it will be clear from the example of Section 5.1, in general we cannot expect strong continuity of the solution with respect to the time. Existence of weak solutions is quite a trivial issue: since the transport equation is a linear PDE, it is sufficient to regularize the vector field and the initial data, obtaining a sequence of smooth solutions to the approximate problems, and then we pass to the limit to get a solution. Theorem 2.3 (Existence of weak solutions). Let b ∈ L∞ ([0, T ] × Rd ; Rd ) with div b ∈ L1loc ([0, T ] × Rd ) and let u ¯ ∈ L∞ (Rd ). Then there exists a weak solution u ∈ ∞ d L ([0, T ] × R ) to (2.1).
28
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
Proof. Let ρ be a convolution kernel on Rd and η be a convolution kernel on R1+d . We define u ¯ = u ¯ ∗ ρ and b = b ∗ η . Let u be the unique solution to the regularized transport equation ( ∂t u + b · ∇u = 0 u(0, ·) = u ¯ . This solution exists and is unique for every > 0, since the vector field b is smooth. From the explicit formula (1.16), for the solution of the transport equation with smooth vector field, we immediately deduce that {u } is equi-bounded in L∞ ([0, T ]×Rd ). Thus we can find a subsequence {uj }j which converges in L∞ ([0, T ] × Rd ) − w∗ to a function u ∈ L∞ ([0, T ] × Rd ). Recalling the weak formulation (2.4) it is immediate to deduce that u is a solution of (2.1). 3. Renormalized solutions We now want to face the uniqueness issue. As we already noticed, the application of the chain-rule that we needed in the computations in Section 2.1 is not always justified. We define here the notion of renormalized solution, which corresponds to the idea of “solution which satisfies the desired chain-rule”. Definition 2.4 (Renormalized solution). Let b : I×Rd → Rd be a locally summable vector field such that div b is locally summable, where I ⊂ R is an interval. Let u ∈ L∞ (I × Rd ) be a solution of the transport equation ∂t u + b · ∇u = 0 . We say that u is a renormalized solution if the equation (2.6)
∂t β(u) + b · ∇β(u) = 0
holds in the sense of distributions in I × Rd for every function β ∈ C 1 (R; R). This property, if satisfied by all bounded weak solutions, can be transferred into a property of the vector field itself. Definition 2.5 (Renormalization property). Let b : I × Rd → Rd be a locally summable vector field such that div b is locally summable, where I ⊂ R is an interval. We say that b has the renormalization property if every bounded solution of the transport equation with vector field b is a renormalized solution. The importance of the renormalization property is summarized in the following theorem, which corresponds to the rough statement “renormalization implies wellposedness”. Theorem 2.6. Let b : [0, T ] × Rd → Rd be a bounded vector field with div b ∈ Define the vector field ˜b :] − ∞, T ] × Rd → Rd according to 0 if t < 0 ˜b(t, x) = (2.7) b(t, x) if 0 ≤ t ≤ T .
L1 ([0, T ]; L∞ (Rd )).
If ˜b has the renormalization property, then bounded solutions of the transport equation (2.1) are unique and stable. By stability we mean the following: let bk and u ¯k be two
3. RENORMALIZED SOLUTIONS
29
smooth approximating sequences converging strongly in L1loc to b and u ¯ respectively, with k¯ uk k∞ uniformly bounded; then the solutions uk of the corresponding transport equations converge strongly in L1loc to the solution u of (2.1). Proof. Uniqueness. Since the transport equation is linear, it is sufficient to show that the only bounded solution to the Cauchy problem ∂t u + b · ∇u = 0 in D0 ([0, T ] × Rd ) u(0, ·) = 0 is u ≡ 0. Extending b to negative times as in the statement of the theorem, we obtain that the following equation is satisfied, for t ∈] − ∞, T ]: ∂t u + ˜b · ∇u = 0 in D0 (] − ∞, T ] × Rd ). u(t, ·) = 0 for t ≤ 0 Since we are assuming that ˜b has the renormalization property we deduce that ∂t u2 + ˜b · ∇u2 = 0 (2.8) in D0 (] − ∞, T ] × Rd ). u(t, ·) = 0 for t ≤ 0 Now fix R > 0 and η > 0 and consider a test function ϕ ∈ Cc∞ (] − ∞, T [×Rd ) such that ϕ = 1 on [0, T − η] × BR (0) and ∂t ϕ(t, x) ≤ −kbk∞ |∇ϕ(t, x)|
(2.9) We define
Z f (t) =
on [0, T ] × Rd .
u2 (t, x)ϕ(t, x) dx .
Rd
For every positive ψ ∈ Cc∞ (] − ∞, T [) we test (2.8) against ψ(t)ϕ(t, x) and we apply Fubini’s theorem to deduce Z T Z T Z h i 0 − f (t)ψ (t) dt = u2 (t, x)ψ(t) ∂t ϕ(t, x) + ˜b(t, x) · ∇ϕ(t, x) dx dt −∞
Rd
−∞
Z
T
+
f (t)div ˜b(t, x)ψ(t) dt
−∞ (2.9)
Z
T
≤ −∞ Z T
≤ −∞
f (t)div ˜b(t, x)ψ(t) dt f (t)kdiv ˜b(t, ·)kL∞ (Rd ) ψ(t) dt .
This means that we have d f (t) ≤ kdiv ˜b(t, ·)kL∞ (Rd ) f (t) dt in the sense of distributions in ] − ∞, T ]. But we know, by (2.8), that f (t) = 0 for t < 0. Using the assumption div b ∈ L1 ([0, T ]; L∞ (Rd )) (which clearly implies div ˜b ∈ L1 (] − ∞, T ]; L∞ (Rd ))) we can apply the Gronwall lemma and deduce that Z f (t) = u2 (t, x)ϕ(t, x) dx = 0 Rd
30
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
for every t ∈] − ∞, T ]. From the arbitrariness of ϕ(t, x) we deduce that u(t, ·) = 0 for every t ∈ [0, T ], as desired. Stability. Arguing as in Theorem 2.3, we easily deduce that, up to subsequences, uk converges in L∞ ([0, T ] × Rd ) − w∗ to a distributional solution u of (2.1). However, by the uniqueness part of the theorem, we know that this solution is unique, and then the whole sequence converges to u. Since bk and uk are both smooth, u2k solves the transport equation with vector field bk and initial data u ¯2k . Arguing as before, u2k must ∞ d ∗ converge in L ([0, T ] × R ) − w to the unique solution of (2.1) with initial data u ¯2 . But by the renormalization property this solution is u2 . Hence we have shown that ∗ ∗ uk * u and u2k * u2 in L∞ ([0, T ] × Rd ) − w∗ , and this eventually implies uk → u strongly in L1loc ([0, T ] × Rd ). Remark 2.7. In the previous theorem the extension to negative times in (2.7) is necessary: it is a way of requiring the strong continuity of the solution at the initial time (see the discussion in Section 2.1 and Theorem 3.1). Another possibility would be to change the definition of renormalized solution, asking that β(u) should satisfy the transport equation and have initial data β(¯ u) (this is the approach of [76] and [78]). The renormalization property for ˜b in general does not follow from the renormalization property for b. Moreover, even if b has the renormalization property in the sense of Definition 2.5, it could exist a solution u which is renormalized, but such that β(u)(0, ·) 6= β(¯ u). An explicit example is given by the oscillatory solution constructed by Depauw [81] and presented in Section 5.1. In this case the vector field b has the renormalization property (in [0, 1] × R2 ) but the extension ˜b has not the renormalization property in ] − ∞, 1] × R2 ; the Cauchy problem for the transport equation with initial data u ¯ = 0 has more than one solution. This example also shows the sharpness of the assumptions of Theorem 2.12: indeed, b ∈ L1loc (]0, 1[; BVloc (R2 ; R2 )) but b 6∈ L1loc ([0, 1[; BVloc (R2 ; R2 )), thus the extension ˜b 6∈ L1loc (] − ∞, 1[; BVloc (R2 ; R2 )). Remark 2.8. The result in Theorem 2.6, together with the renormalization results we are going to prove in the rest of this chapter, extends to transport equations with a linear right hand side of order zero in u of the form ∂t u + b · ∇u = cu , d L1 ([0, T ]; L∞ loc (R )).
with c ∈ The only modification in the proof consists in taking into account an additional term in the Gronwall argument. In particular, choosing c = −div b, we are able to translate all the forthcoming well-posedness results for the transport equation into well-posedeness results for the continuity equation. Remark 2.9. When dealing with the continuity p equation ∂t u +1 div (bu) = 0 we can consider the renormalization function β(z) = 1 + (z + )2 − 1 ∈ C (R). Integrating formally over Rd we obtain Z Z d β(u(t, x)) dx = div b(t, x) β(u(t, x)) − u(t, x)β 0 (u(t, x)) dx . dt Rd Rd Since for this choice of β we have β(z) − zβ 0 (z) ≤ 0, it turns out that it is enough to have a control on the negative part of the divergence: we only need to assume [div b]− ∈ L1 ([0, T ]; L∞ (Rd )).
4. THE DIPERNA–LIONS REGULARIZATION SCHEME
31
Since in all the forthcoming renormalization results no regularity (but just integrability) is required with respect to the time, it is clear that the zero extension for negative times remains in the same class of regularity initially assumed for the vector field. We will deal with vector fields defined on I × Rd , where I ⊂ R is a generic interval of times, since in order to treat the well-posedness problem for t ∈ [0, T ] we need to show the renormalization property for the extended vector field, defined on the time interval I =] − ∞, T ]. In the following we are going to illustrate various cases in which the renormalization property holds: in all these cases the well-posedness for the transport equation and for the continuity equation presented in Theorem 2.6 will then hold. 4. The DiPerna–Lions regularization scheme As we remarked in the initial discussion of Section 2.1, the main obstruction to the renormalization property is the lack of regularity of the solution u. It is therefore not surprising that the main technique used to prove this property is a regularization procedure. This idea goes back to the work by DiPerna and Lions [83], where it was used to show the renormalization property for Sobolev vector fields. The link between renormalization and more general regularization procedures will be also investigated in the next chapter: see in particular Theorem 3.1 and Remark 3.6. Let us fix an even convolution kernel ρ in Rd . We start from the transport equation ∂t u + b · ∇u = 0 and we convolve it by ρ . We define u = u ∗ ρ and we notice that we can let the convolution act on u in b · ∇u only at the price of an error term: (2.10) ∂t u + b · ∇u = b · ∇u − b · ∇u ∗ ρ . We define the commutator r as the error term in the right hand side of the previous equality: (2.11) r = b · ∇u − b · ∇u ∗ ρ . We call this term commutator because it measures the difference in exchanging the two operations of convolution and of differentiation in the direction of b. Now notice that the function u is smooth with respect to the spatial variable. Moreover we have the identity ∂t u = − b · ∇u ∗ ρ , hence, for every > 0 fixed, the function u has Sobolev regularity in space-time. This implies that, if we multiply both sides of (2.10) by β 0 (u ), we can apply Stampacchia’s chain-rule for Sobolev maps (see for instance Section 4.2.2 of [90]) to get β 0 (u ) ∂t u + b · ∇u = ∂t β(u ) + b · ∇β(u ) . Thus, for every > 0 we have the equality (2.12)
∂t β(u ) + b · ∇β(u ) = r β 0 (u ) .
Now we would like to pass to the limit as → 0 in order to recover the renormalization property. The left hand side converges in the sense of distributions to the left hand side of (2.6). Thus we need to show the convergence to zero of the quantity r β 0 (u ).
32
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
Notice that r converges to zero in the sense of distributions, without any regularity assumption on b. The sequence β 0 (u ) is equi-bounded in L∞ and converges L d -a.e. for every t ∈ I, but this does not allow to deduce the convergence of the product r β 0 (u ). A first attempt would be to show the strong convergence to zero of the commutator r . This is precisely the result that DiPerna and Lions obtained in [83] under a Sobolev regularity assumption.
5. Vector fields with Sobolev regularity Theorem 2.10 (DiPerna–Lions). Let b be a bounded vector field belonging to 1,p L1loc (I; Wloc (Rd ; Rd )), where I ⊂ R is an interval. Then b has the renormalization property. Proof. The theorem immediately follows from the discussion in the previous section and from the following proposition, in which strong convergence to zero of the commutator is proved. Proposition 2.11 (Strong convergence of the commutator). Let b be a bounded 1,p (Rd ; Rd )), where I ⊂ R is an interval, and let vector field belonging to L1loc (I; Wloc ∞ d u ∈ Lloc (I × R ). Define the commutator r as in (2.11). Then r → 0 strongly in L1loc (I × Rd ). Proof. Playing with the definitions of b·∇u and of the convolution of a distribution and a smooth function it is not difficult to show the following explicit expression of the commutator: Z b(t, x) − b(t, x − z) (2.13) r (t, x) = u(t, x − z) · ∇ρ(z) dz + udiv b ∗ ρ . Rd We recall the following property, which in fact characterizes functions in Sobolev spaces 1,1 we have (see Theorem 2.1.6 in [130]): for every f ∈ Wloc (2.14)
f (x + z) − f (x) → ∇f (x)z
strongly in L1loc as → 0,
that is, the difference quotients converge strongly to the derivative. Using (2.14) and the strong convergence of translations in Lp in (2.13) we obtain that r converges strongly in L1loc (I × Rd ) to Z u(t, x) ∇b(t, x)z · ∇ρ(z) dz + udiv b . Rd
The elementary identity Z zi Rd
∂ρ(z) dz = −δij ∂zj
immediately shows that the limit is zero. This concludes the proof.
6. VECTOR FIELDS WITH BOUNDED VARIATION
33
6. Vector fields with bounded variation In this section we present the main lines of the argument by Ambrosio [8] for the renormalization for vector fields with bounded variation. The main point which makes the difference with respect to the Sobolev case is the lack of strong convergence of the difference quotients (compare with (2.14)). The proof will involve a splitting of the difference quotients and an anisotropic regularization, based on a local selection of a “bad” direction given by Alberti’s Rank-one theorem. Theorem 2.12 (Ambrosio). Let b be a bounded vector field belonging to such that div b L d for L 1 -a.e. t ∈ I. Then b has the renormalization property. L1loc (I; BVloc (Rd ; Rd )),
We intend to give a reasonably detailed description of the proof of this result, mostly trying to enlighten the main ideas. We will present an approach which is slightly simpler than the one of the original paper [8]. We also refer to [9], [10] and [78] for an account of the proof of Ambrosio’s theorem. This theorem has been applied to the study of various nonlinear PDEs: for instance the Keyfitz and Kranzer system ([15] and [11]) and the semigeostrophic equation ([72], [73] and [31]). We start by introducing some notation. For the basic terminology and facts about BV functions we refer to Appendix 3. For every t ∈ I we perform the decomposition of the spatial derivative of b(t, ·) (which is, by definition, a matrix-valued measure in Rd ) in absolutely continuous and singular part as follows: Db(t, ·) = Da b(t, ·) + Ds b(t, ·) . Moreover, we denote by |Db|, |Da b| and |Ds b| the measures obtained by integration of |D(t, ·)|, |Da b(t, ·)| and |Ds b(t, ·)| with respect to the time variable, that is Z Z Z ϕ(t, x) d|Db|(t, x) = ϕ(t, x) d|Db(t, ·)|(x) dt , I×Rd I Rd Z Z Z ϕ(t, x) d|Dσ b|(t, x) = ϕ(t, x) d|Dσ b(t, ·)|(x) dt , σ = a, s , I×Rd
I
Rd
for every function ϕ ∈ Cc (I × Rd ). 6.1. Difference quotients of BV functions. As we already observed, the strong convergence of the difference quotients in (2.14) characterizes functions in Sobolev spaces. Hence a first step in the proof of Theorem 2.12 is a description of the behaviour of the difference quotients for a function with bounded variation. For any function f ∈ BVloc and any z ∈ Rd with |z| ≤ we have the classical L1 estimate of the increments Z (2.15) |f (x + z) − f (x)| dx ≤ |Dz f |(K ) for any compact set K ⊂ Rd , K
where Dz f = (Df )z denotes the component of Df in the direction z and K is the open -neighbourhood of K. Moreover, we can give a canonical decomposition of the difference quotients in order to analyse independently the behaviour of the absolutely continuous part and of the
34
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
singular part. Let us start from the case d = 1. If µ is an Rm -valued measure in R with locally finite variation, then by Jensen’s inequality the functions µ([t, t + ]) , t∈R µ b (t) = satisfy Z |b µ (t)| dt ≤ |µ|(K )
for every compact set K ⊂ R.
K
A simple density argument then shows that µ b converge in L1loc (R) to the density of µ 1 1 with respect to L whenever µ L . If f ∈ BVloc (R; Rm ) and > 0 we know that f (x + ) − f (x) Df ([x, x + ]) Da f ([x, x + ]) Ds f ([x, x + ]) = = + 1 for L -a.e. x ∈ R, the exceptional set possibly depending on . In this way we have canonically split the difference quotient of f as the sum of two functions, one strongly convergent to the absolutely continuous part of the derivative Da f , and the other one having an L1 norm on any compact set K asymptotically smaller than |Ds f |(K). This argument can be carried on also in the multi-dimensional case with the help of the slicing theory of BV functions. If we fix the direction z ∈ Rd , we obtain that the difference quotients can be canonically split into two parts as follows b(t, x) − b(t, x − z) = b1,z (t, x) + b2,z (t, x) , (2.16) with b1,z (t, x) → Da b(t, x) z
(2.17)
strongly in L1loc (Rd )
and Z (2.18) lim sup →0
|b2,z (t, x)| dx ≤ |Ds b(t, x) z|(K)
for every compact set K ⊂ Rd .
K
6.2. Isotropic estimate and the defect measure σ. We proceed with the DiPerna–Lions regularization scheme illustrated in Section 2.4. We arrive at the expression (2.13) for the commutator and we insert in it the decomposition (2.16) of the difference quotients, getting r = b · ∇u − b · ∇u ∗ ρ Z b(t, x) − b(t, x − z) = u(t, x − z) · ∇ρ(z) dz + udiv b ∗ ρ d ZR (2.19) = u(t, x − z)b1,z (t, x) · ∇ρ(z) dz + udiv b ∗ ρ Rd Z + (2.20) u(t, x − z)b2,z (t, x) · ∇ρ(z) dz . Rd
Arguing as in Section 2.5, applying (2.17) we deduce that (2.19) converges strongly to Z u(t, x) Da b(t, x)z · ∇ρ(z) dz + u(t, x) div b(t, x) d R = − u(t, x) trace Da b(t, x) + u(t, x) div b(t, x) = 0 ,
6. VECTOR FIELDS WITH BOUNDED VARIATION
35
since trace Da b(t, x) is the absolutely continuous part of the divergence, which however coincides with the whole divergence, because we are assuming that div b L d . For the integral in (2.20) we use the estimate (2.18) for b2,z to deduce the isotropic estimate Z Z (2.21) lim sup |r | dtdx ≤ kuk∞ I(ρ)|Ds b|(J × K) , →0
J
K
Rd
for every compact set K ⊂ and every interval J ⊂ I, where we have defined the isotropic energy of the convolution kernel ρ to be Z |z| |∇ρ(z)| dz . I(ρ) = Rd
From the isotropic estimate (2.21) and the equi-boundedness in L∞ (I × Rd ) of {β 0 (u )} we deduce that the sequence {r β 0 (u )} has limit points in the sense of measures. Hence we can introduce the defect measure σ = ∂t β(u) + b · ∇β(u)
(2.22)
and we deduce from (2.21) that σ is a locally finite measure satisfying (2.23)
|σ| ≤ kβ 0 k∞ kuk∞ I(ρ)|Ds b|
as measures on I × Rd .
In particular, this shows that σ is singular with respect to L 1 ⊗ L d . We can interpret (2.22) by saying that, up to now, we have shown the renormalization property up to an error term which is a singular measure. 6.3. Anisotropic estimate. We now give another estimate for the local norm of the commutator r . In this estimate we do not take any advantage of the cancellations, but we introduce an anisotropic term related to the derivative of the vector field b which will be useful in the final estimate. We denote by Mt the matrix which appears in the polar decomposition of Db(t, ·), that is the (d × d)-matrix defined |Db(t, ·)|-a.e. by the identity Db(t, ·) = Mt (·)|Db(t, ·)|. For every z ∈ Rd and for L 1 -a.e. t ∈ I we have the identity (2.24)
Dz b(t, ·) · ∇ρ(z) = Mt (·)z · ∇ρ(z) |Db(t, ·)|
as measures on Rd .
Going back to (2.13) and using (2.24) and the L1 estimate (2.15) we obtain Z Z Z Z (2.25) lim sup |r | dtdx ≤ Λ(Mt (x), ρ) d|Db|(t, x) + d|Da b|(J × K) →0
J
K
J
K
Rd
for every compact set K ⊂ and every interval J ⊂ I. Here, for every d × d matrix N and for every convolution kernel ρ, we have defined the anisotropic energy of ρ as Z Λ(N, ρ) = |N z · ∇ρ(z)| dz . Rd
This implies that the defect measure σ satisfies (2.26) |σ| ≤ kβ 0 k∞ kuk∞ Λ(Mt (·), ρ)|Db| + d|Da b|
as measures on I × Rd .
36
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
6.4. Combination of the two estimates. From (2.23) and (2.26) we obtain that (2.27)
|σ| ≤ kβ 0 k∞ kuk∞ Λ(Mt (·), ρ)|Ds b|
as measures on I × Rd .
This can be deduced from the following general fact: if σ, µ1 and µ2 are positive measures such that σ ≤ µ1 + µ2 and σ ⊥ µ2 , then σ ≤ µ1 . Now we observe that (2.27) holds for every even convolution kernel ρ. The important point here is that the defect measure σ does not depend on ρ. Our aim is to show that σ = 0, and this will be achieved optimizing this estimate. Since the estimate has a local nature, this optimization procedure is, in a certain sense, equivalent to vary the regularizing kernel in t and x. We claim that 0 (2.28) |σ| ≤ kβ k∞ kuk∞ inf Λ(Mt (·), ρ) |Ds b| , ρ
where the infimum is taken over the set Z ∞ d K = ρ ∈ Cc (R ), such that ρ ≥ 0 is even and
ρ=1 .
Rd
We first notice that (2.27) gives that σ f (t, x) such that
|Ds b|,
hence there exists a Borel function
σ = f (t, x)|Ds b| .
(2.29) Hence we can rewrite (2.27) as (2.30)
|f (t, x)| ≤ kβ 0 k∞ kuk∞ Λ(Mt (x), ρ)
for |Ds b|-a.e. (t, x)
for every fixed ρ. Notice however that the set where the inequality (2.30) fails could in principle depend on ρ. This is not a problem as soon as we infimize on a countable set of kernels K0 ⊂ K: we can deduce 0 (2.31) |f (t, x)| ≤ kβ k∞ kuk∞ inf 0 Λ(Mt (x), ρ) for |Ds b|-a.e. (t, x). ρ∈K
However we notice that, for every fixed matrix M , the map ρ 7→ Λ(M, ρ) is continuous with respect to the W 1,1 topology. Choosing K0 ⊂ K countable and dense with respect to the W 1,1 topology we obtain that the infimum over K0 coincides with the infimum over K. This implies (2.28). 6.5. Local optimization of the convolution kernel. In order to show that σ = 0 and to obtain the renormalization property we only need to show that the infimum in (2.28) is actually equal to zero, under the assumption that b has bounded variation and div b is absolutely continuous with respect to L d . The original proof in [8] is based on Alberti’s Rank-one theorem (see Theorem A.6) and on an anisotropic regularization procedure (first introduced by Bouchut [36] in the context of the Vlasov equation with BV field and subsequently exploited by Colombini and Lerner [60] to treat the transport equation with conormal-BV vector field). We present in the next subsection a simplified version of the proof, based on a lemma due to Alberti, in which the infimum in (2.28) is computed.
6. VECTOR FIELDS WITH BOUNDED VARIATION
37
The anisotropic regularization procedure of Bouchut is based on the following heuristic remark. Suppose that the vector field b depends “nicely” on a group of variables x1 and “badly” on another group x2 . Then it is more efficient to regularize quickly along the direction x2 and slowly along the direction x1 , in order to rule out faster the bad dependence in x2 . This can be achieved using anisotropic convolution kernels, which are asymptotically very close to the characteristic function of a very thin rectangle, whose long side is along the x1 direction. This is applied by Colombini and Lerner [60] to deal with the case of conormal-BV vector fields, in which the derivative of b is a measure along one direction and an L1 function along the other d − 1 directions. Then they regularize much faster in the direction along which the derivative is a measure, according to Bouchut’s scheme. In [8] the selection of the two directions is based on Alberti’s Rank-one theorem (see Theorem A.6). This result allows to select, for |Ds b(t, ·)|-a.e. x ∈ Rd , two unit vectors ξt and ηt in such a way that Ds b(t, ·) = ξt ⊗ ηt |Ds b(t, ·)| . This can be intepreted as an asymptotic dependence of b(t, ·) on the direction ηt only, hence the shape of the anisotropic kernel is carefully chosen to be very thin in this direction. 6.6. Alberti’s lemma. In this section we describe Alberti’s lemma, which allows to compute exactly the infimum Z Λ(M ) = inf Λ(M, ρ) = inf |M z · ∇ρ(z)|dz, ρ∈K
ρ∈K Rd
for every d × d matrix M . The proof is again related to an anisotropic regularization: the basic idea is in some sense to take a convolution kernel which is concentrated on a very long tube made of trajectories of the ordinary differential equation γ˙ = M γ. Lemma 2.13. For every d × d matrix M we have Λ(M ) = |trace (M )| . We observe first that this result allows to conclude the proof of Theorem 2.12. Indeed, from the assumption div b L d for L 1 -a.e. t ∈ I, it follows that trace (Mt (·))|Ds b(t, ·)| = 0
for L 1 -a.e. t ∈ I.
This means that inf Λ(Mt (·), ρ) = 0
ρ∈K
for |Ds b|-a.e. (t, x) ∈ I × Rd . Going back to (2.28) we deduce that σ = 0, hence we have proved the renormalization result of Theorem 2.12.
38
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
Proof of Lemma 2.13. We first observe that the inequality Λ(M ) ≥ |trace (M )| is trivial. Indeed for every admissible kernel ρ we have Z Z |M z · ∇ρ(z)| dz ≥ M z · ∇ρ(z) dz d d R R Z div (M z)ρ(z) dz = − Rd Z ρ(z) dz = |trace (M )| . = − trace (M ) Rd
We pass to the opposite inequality. Since M z · ∇ρ(z) = div (M z ρ(z)) − trace (M ) ρ(z) and ρ ≥ 0, we immediately obtain Z Z |M z · ∇ρ(z)| dz ≤ |div (M z ρ(z))| + |trace (M )| ρ(z) dz d Rd ZR = |div (M z ρ(z))| dz + |trace (M )| . Rd
Then it suffices to show that Z (2.32)
inf
ρ∈K Rd
|div (M z ρ(z))| dz = 0 .
Step 1. We preliminarily show that Z (2.33) inf |div (M z µ)| = 0 , µ
Rd
where the infimum is now taken over all probability measures µ on Rd . With the integral we intend the total variation of the distribution div (M z µ), i.e. n o sup hM z µ, ∇ϕi : ϕ ∈ Cc∞ (Rd ) , kϕk∞ ≤ 1 . In particular, the total variation is finite if and only if div (M z µ) is a finite measure, and if this is the case it coincides with kdiv (M z µ)kM(Rd ) . Without loss of generality, we can assume M 6= 0. Now let us consider µ = ξH 1 Γ, where Γ is the support of a nonconstant integral line γ : [0, T ] → Rd of the vector field M z, i.e. γ satisfies γ(t) ˙ = M γ(t) for every t ∈ [0, T ]. The density ξ is of the form ξ(z) = #{γ −1 (z)}
α , |M z|
where α is the normalization constant Z T Z #{γ −1 (z)} 1 −1 1 α = dH (z) = |γ(t)| ˙ dt = T . |M z| |M γ(t)| Γ 0
6. VECTOR FIELDS WITH BOUNDED VARIATION
39
Since Z hdiv (M z µ), ϕi = −hM z µ, ∇ϕi = − M z · ∇ϕ(z) dµ(z) Rd Z #{γ −1 (z)} dH 1 (z) = − M z · ∇ϕ(z) T |M z| Γ Z 1 T =− γ(t) ˙ · ∇ϕ(γ(t)) dt T 0 1 = − ϕ(γ(T )) − ϕ(γ(0)) T for every ϕ ∈ Cc∞ (Rd ), we get div (M z µ) = −
1 δγ(T ) − δγ(0) T
in D0 (Rd ).
Hence Z |div (M z µ)| ≤ Rd
2 T
for any such µ. Then (2.33) follows immediately. Indeed, taken any maximal integral line γ : [0, +∞[→ Rd , it is enough to consider its restriction to [0, T ] and then to use the arbitrariness of the choice of T . Step 2. Now we show the validity of (2.32). For every x ∈ Rd we consider the maximal solution γx : [0, +∞[→ Rd of the problem ( γ˙ x (t) = M γx (t) γx (0) = x . It exists for every initial point x ∈ Rd and is explicitly given by the expression γx (t) = etM x. We define A = x ∈ Rd : γx is not a constant . Notice that A is a nonempty open set which is symmetrical with respect to the origin. R Now fix an even and positive function θ ∈ Cc∞ (Rd ) with Rd θ = 1 and such that spt θ ⊂ A. We set Z T ν = µTx θ(x) dx , Rd
with µTx defined as before, starting from γx |[0,T ] . Let us check that ν T is in fact a smooth function. If ϕ ∈ Cc (Rd ) we have hµTx , ϕi
1 = T
Z
T
ϕ(γx (t)) dt . 0
40
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
Hence we deduce the following expression for ν T : Z Z Z 1 T T T hµx , ϕiθ(x) dx = ϕ(etM x)θ(x) dxdt hν , ϕi = T 0 Rd Rd Z Z 1 T = ϕ(y)θ(e−tM y) det(e−tM ) dydt T 0 Rd ! Z Z 1 T = θ(e−tM y) det(e−tM )dt ϕ(y)dy . T 0 Rd It is then clear that ν T is a smooth function which satisfies all the constraints of the problem (2.32). Moreover Z divz (M z µTx )θ(x) dx , divz (M z ν T (z)) = Rd
where the divergence inside the integral is in the sense of distributions, for every fixed value of x. Eventually we get Z 2 |divz (M z ν T (z))| dz ≤ . T d R From the arbitrariness of T we obtain (2.32).
7. Other various renormalization results We now briefly indicate some other renormalization results for vector fields with some particular structure. We will not enter into the proofs of such results but we will rather give some basic references on each of them. 7.1. Conditions on the symmetric part of the derivative. In [50] Capuzzo Dolcetta and Perthame show, among other things, that the renormalization property holds for vector fields b such that the symmetric derivative d×d 1 Eb = Db + t Db ∈ D0 (I × Rd ) 2 is absolutely continuous with respect to the Lebesgue measure. Their proof is an adaptation of DiPerna–Lions’ one: if the convolution kernel ρ used in the regularization procedure is chosen to be radial then it is simple to see that, for every direction z ∈ Rd , in order to control the commutator it is not necessary to control the whole difference quotient in the direction z, but only its scalar product with z itself. This control is precisely the one ensured by the condition on the symmetric derivative. It would be interesting to adapt this proof to show the renormalization property for vector fields with bounded deformation, i.e. such that the symmetric part of the derivative is a measure (see Appendix 3 for a more precise definition of the space BD). However, as we have just observed, conditions on the symmetric part of the derivative seem exploitable only through the use of radial convolution kernels; hence the whole anisotropic regularization technique is apparently not disposable in this context. A result in this direction has been obtained in a work in collaboration with Ambrosio and Maniglia [14]: we are able to treat the case of special vector fields with bounded
8. NEARLY INCOMPRESSIBLE VECTOR FIELDS
41
deformation, which satisfy by assumption the condition that the singular part of the symmetric derivative is concentrated on a set of codimension one. The singular part is ruled out using a chain-rule formula for normal traces of vector fields with measure divergence. See also [25], [55] and [56] for general results on this notion of trace. However the Cantor part of the derivative is not treatable with this kind of techniques. 7.2. Partial regularity of the vector field. Looking at the results by DiPerna and Lions and by Ambrosio it is apparent that the regularity needed on b in the dependence on the spatial variable (Sobolev or BV ) is much greater than the one needed in the dependence on the time variable (just summability is requested). If we look at the vector field B(t, x) = (1, b(t, x)) in space-time, for which the transport equation reads B(t, x) · ∇t,x u(t, x) = 0 , we see that B has a kind of “split structure”, in which the first component depends fairly on the first variable, and the second component depends fairly on the second variable, but very wildly on the first one. It is then natural to conjecture (see Remark 3.8 of [8]) that the renormalization property should hold for vector fields which admit a splitting of the type B = b1 (x1 ), b2 (x1 , x2 ) , x1 , b1 ∈ Rd1 , x2 , b2 ∈ Rd2 , under an assumption of good dependence of b1 on x1 and of b2 on x2 . Some results in this direction are available in [103], [104] and [105], in which the partial regularity is of Sobolev or BV type. The proof relies again on an anisotropic convolution: the idea is that we have to regularize faster along the x1 direction, because b2 has no regularity in that direction. It would be interesting to understand if results of this kind can hold if we admit also an x2 dependence of the b1 component. 7.3. The two-dimensional case. Consider a bounded autonomous vector field b : R2 → R2 with div b = 0. Then it is possible to find a Lipschitz function H : R2 → R (called the Hamiltonian) such that b = ∇⊥ H. Heuristically we expect that the evolution is split on the level lines of H, since the value of H is conserved under the flow. Hence we expect a kind of dimensional reduction of the problem, which we could try to solve “line by line”. However, in this very general setting, the well-posedness does not hold: in a joint work with Alberti and Bianchini [5] we construct counterexamples to the uniqueness for bounded autonomous planar divergence-free vector fields. For the uniqueness to hold we need to require a weak Sard property of the Hamiltonian (see (4.11) and Theorem 4.8), under which we are able to implement the idea of the splitting on the level curves of H. This result has been proved in the same paper [5] and will be presented in Chapter 4, where we will also indicate more references on the problem. 8. Nearly incompressible vector fields The aim of this section is to develop a well-posedness theory for vector fields in which the usual assumption of boundedness of the divergence is replaced by a control of the Jacobian, or by the existence of a solution of the continuity equation which is bounded away from zero and infinity. This is particularly important in view of the applications, for instance to the Keyfitz and Kranzer system (see [15] and [11]). A systematic
42
2. RENORMALIZED SOLUTIONS AND WELL-POSEDNESS OF THE PDE
treatement of this topic is done in [76]. Since the assumption on the divergence is necessary in order to give a distributional meaning to the transport equation (see Section 2.2), we deal here with the equation in continuity form. We will consider again the case of nearly incompressible vector fields, from the ODE viewpoint, in Sections 6.6 and 6.7. Definition 2.14 (Nearly incompressible vector fields). We say that a bounded vector field b : I × Rd → Rd is nearly incompressible if there exist a function ρ ∈ L∞ (I × Rd ) and a constant C > 0 such that (2.34)
0
0, such that the identity (3.17)
∂t ρ + div (bρ) = 0
holds in the sense of distributions in [0, T ] × Rd . Theorem 3.14. Let b ∈ L∞ ([0, T ] × Rd ; Rd ) be a nearly incompressible vector field, and fix an associated function ρ ∈ C([0, T ]; L∞ (Rd ) − w∗ ) as in Definition 2.14. We define the Banach space F and its norm k · kF as in (3.2)–(3.3). Let F 1 ⊂ F be the closure of n o ρϕ : ϕ ∈ C ∞ ([0, T ] × Rd ) with compact support in x with respect to k · kF . Then the following statements are equivalent: (i) For every u0 ∈ L2 and every f ∈ L2 there exists a solution u ∈ F 1 to the Cauchy problem ∂t u + div (bu) = f u ∈ F 1. u(0, ·) = u0 , (ii) There is uniqueness for weak solutions in C([0, T ]; L2 (Rd )−w) for the backward dual Cauchy problem starting from T ; i.e., the only function ρv belonging to C([0, T ]; L2 (Rd ) − w) which solves ∂t (ρv) + div (bρv) = 0 ρ(T, ·)v(T, ·) = 0 is ρv ≡ 0. Remark 3.15. In this context, the equation ∂t (ρv) + div (bρv) = 0 is dual to the equation ∂t u + div (bu) = 0, since we can write (formally, since it is not possible to give a meaning to the product b · ∇v without a condition of absolute continuity of div b) ∂t (ρv) + div (bρv) = ρ ∂t v + b · ∇v . Proof of Theorem 3.14. The proof is very close to that of Theorem 3.8; thus we shall sometimes omit the technical details. Step 1. An energy estimate in F 1 . We preliminarily prove that for every u ∈ F 1 the following estimate holds (C is the constant in (2.34)): √ (3.18) kukBt (L2x ) ≤ Cku(0, ·)kL2x + C T k∂t u + div (bu)kL2t,x .
56
3. AN ABSTRACT CHARACTERIZATION OF THE RENORMALIZATION PROPERTY
Fix a smooth function ϕ with compact support in Rd , and define f = ∂t (ρϕ) + div (bρϕ) = ρ(∂t ϕ + b · ∇ϕ) (use the Leibniz rule (3.13) and formula (3.17)). We deduce with the same argument that 2ϕf = ρ(∂t ϕ2 + b · ∇ϕ2 ) = ∂t (ρϕ2 ) + div (bρϕ2 ) . Thus, we get the following estimate in the sense of distributions in [0, T ]: Z Z d 2 ϕ(t, x)f (t, x) dx ρ(t, x)ϕ(t, x) dx = 2 dt Rd Rd ≤ 2kf (t, ·)kL2x kϕ(t, ·)kL2x Z 1/2 √ 2 ρ(t, x)ϕ(t, x) dx . ≤ 2 Ckf (t, ·)kL2x Rd
By integration with respect to time this implies Z 1/2 Z 1/2 √ Z t ρ(t, x)ϕ(t, x)2 dx ≤ ρ(0, x)ϕ(0, x)2 dx + C kf (s, ·)kL2x ds. Rd
Rd
0
Using (2.34) we deduce √ √ 1 √ kρ(t, ·)ϕ(t, ·)kL2x ≤ Ckρ(0, ·)ϕ(0, ·)kL2x + C C and thus (3.19)
Z
t
0
kf (s, ·)kL2x ds,
√ kρ(t, ·)ϕ(t, ·)kL2x ≤ Ckρ(0, ·)ϕ(0, ·)kL2x + C T k∂t (ρϕ) + div (bρϕ)kL2t,x .
But by definition of F 1 , the validity of (3.19) for every smooth function ϕ with compact support in x implies the validity of (3.18) for every function u ∈ F 1 . Step 2. The operator A1 . We define the linear operator 1
F 1 → L2 (Rd ) × L2 ([0, T ] × Rd ),
A : u
7→
u(0, ·), ∂t u + div (bu) .
It is immediate to see that the operator A1 is bounded. Using the energy estimate (3.18) ˜ 1 uk, and therefore that A1 is injective it is also immediate to check that kukF ≤ CkA with closed image. Applying Lemma 3.10(ii) we obtain that the adjoint operator (A1 )∗ : L2 (Rd ) × L2 ([0, T ] × Rd ) → (F 1 )∗ is surjective. The adjoint operator is characterized by the identity
(3.20)
h(A1 )∗ (v0 , v), ui = h(v0 , v), A1 ui Z Z = v0 u(0, ·) dx + Rd
L2 (Rd )
L2 ([0, T ]
v ∂t u + div (bu) dtdx
[0,T ]×Rd
Rd )
for (v0 , v) ∈ × × and u ∈ F 1 . Step 3. Proof of the equivalence of the two statements. Statement (i) (existence of solutions in F 1 ) is the surjectivity of the operator A1 , which is equivalent (applying Lemma 3.10(i) and using the surjectivity of (A1 )∗ proved in Step 2) to the
3. NEARLY INCOMPRESSIBLE VECTOR FIELDS
57
injectivity of (A1 )∗ . But recalling the characterization (3.20) and the definition of the space F 1 , we see that the injectivity of (A1 )∗ is equivalent to the following implication for v0 ∈ L2 (Rd ) and v ∈ L2 ([0, T ] × Rd ): (3.21) R R v0 = 0 [0,T ]×Rd ∂t (ρϕ) + div (bρϕ) v dtdx + Rd ρ(0, ·)ϕ(0, ·)v0 dx = 0 =⇒ . v=0 ∀ϕ ∈ C ∞ ([0, T ] × Rd ) with compact support in x Arguing as in Step 3 of the proof of Theorem 3.8 we obtain that the following two properties are equivalent: • For every ϕ ∈ C ∞ ([0, T ] × Rd ) with compact support in x we have Z Z ∂t (ρϕ) + div (bρϕ) v dtdx + ρ(0, ·)ϕ(0, ·)v0 dx = 0. [0,T ]×Rd
Rd
• ρv ∈ C([0, T ]; L2 (Rd ) − w), ρ(0, ·)v0 = ρ(0, ·)v(0, ·), and ρv is a weak solution of the backward dual Cauchy problem ∂t (ρv) + div (bρv) = 0 ρ(T, ·)v(T, ·) = 0. Then we deduce that implication (3.21) is equivalent to statement (ii), and this concludes the proof of the theorem.
CHAPTER 4
Well-posedness in the two-dimensional case In this chapter we describe some well-posedness results that are available in the two-dimensional case. Due to the special structure of the problem, which admits a Hamiltonian function conserved (at least formally) by the flow, the assumptions needed for the uniqueness are dramatically weaker than those of Chapter 2. We start in the first section by presenting some standard considerations and collecting some results available in the literature; we also mention a first result obtained with Colombini and Rauch [57] which goes beyond the divergence-free assumption. The rest of the chapter is devoted to the presentation of a work in progress in collaboration with Alberti and Bianchini [5], in which sharp well-posedness results in the two-dimensional case are obtained. We present here just the basic case of a bounded divergence-free vector field, while some variations are possible. The uniqueness holds under an additional assumption: we must require the weak Sard property (4.11), which turns out to be necessary, in view of the counterexamples contained in [5]. 1. Bounded planar divergence-free vector fields Let us consider an autonomous vector field b ∈ L∞ (R2 ; R2 ) in the plane such that div b = 0. It is well-known that in this situation it is possible to find a Hamiltonian function H ∈ Lip(R2 ) such that ∂H(x) ∂H(x) ⊥ , for L 2 -a.e. x ∈ R2 . (4.1) b(x) = ∇ H(x) = − ∂x2 ∂x1 The starting point for all the two-dimensional well-posedness results is the heuristic remark that the value of the Hamiltonian is constant on the trajectories. Indeed, if γ(t) ˙ = b(γ(t)), then we can compute d H(γ(t)) = ∇H(γ(t)) · γ(t) ˙ = ∇H(γ(t)) · b(γ(t)) = ∇H(γ(t)) · ∇⊥ H(γ(t)) = 0 . dt This means that the trajectories “follow” the level sets of the Hamiltonian. Heuristically, one can try to implement the following strategy: (a) Localize the equation to each level set, thanks to the fact that the level sets are invariant under the action of the flow; (b) Understand the structure of the level sets, trying to prove that generically they are “one-dimensional sets”; (c) See the equation on each level set as a one-dimensional problem and show uniqueness for it; 59
60
4. WELL-POSEDNESS IN THE TWO-DIMENSIONAL CASE
(d) Deduce uniqueness for the problem in R2 from the uniqueness of all the problems on the level sets. Since we can hope for uniqueness on the level sets under quite general hypotheses, the reduced equation being one-dimensional, we expect stronger well-posedness results in this case: it is natural to imagine that no regularity of b (in terms of weak derivatives) would be needed. We first indicate the essential literature on this subject. Previous results by Bouchut and Desvillettes [38], Hauray [99] and Colombini and Lerner ([58] and [60]) show that uniqueness holds for the transport equation relative to an autonomous bounded divergence-free vector field, under the following additional condition on the local direction of the vector field: there exists an open set Ω ⊂ R2 such that H 1 (R2 \ Ω) = 0 and for every x ∈ Ω the following holds: (4.2)
there exist ξ ∈ S1 , α > 0 and > 0 such that, for L 2 -a.e. y ∈ B (x), we have b(y) · ξ ≥ α.
The validity of this condition permits a local change of variable, which straightens the level sets of the Hamiltonian, thus reducing the equation to a one-dimensional problem (the second spatial variable appears as a parameter in the equation after the change of variable). A first extension to the non-divergence-free case is due to Colombini and Rauch [63]: they are able to show that the uniqueness holds in the case of autonomous bounded vector fields with bounded divergence for which there exists a positive Lipschitz function θ, bounded and bounded away from zero, such that (4.3)
div (θb) = 0 .
In [63] it is also conjectured that this hypothesis on the existence of the function θ could be removed. In a subsequent paper in collaboration with Colombini and Rauch [57] we show that this is the case: in fact we are able to prove the following result. Theorem 4.1. Assume that b ∈ L∞ (R2 ; R2 ) and that div b ∈ L∞ (R2 ). Assume that there exists an open set Ω ⊂ R2 such that H 1 (R2 \ Ω) = 0 and that the condition in (4.2) holds for every x ∈ Ω. Then we have uniqueness in L∞ ([0, T ] × R2 ) for the Cauchy problem for the transport equation. However, the meaning of condition (4.2) is not completely clear: while in the stationary problem it just expresses the fact the surface on which we consider the initial data is noncharacteristic, in the time-space problem it is a kind of local regularity of the direction of b. In particular, condition (4.2) prevents the existence of “too many” zeros of the vector field. In the rest of this chapter we present a more recent result, still in progress, in collaboration with Alberti and Bianchini [5]. In this paper the strategy is a bit different: we do not perform a local change of variable according to the Hamiltonian, but we rather split the equation on the level sets of the Hamiltonian, using the coarea formula. Then we would like to look at the equation level set by level set. It turns out that, where ∇H 6= 0, the level sets are in fact nice rectifiable curves, and this will allow to consider the PDE in the parametrization. The interesting point is that, in order to separate the evolution in {∇H = 0} from the evolution in {∇H 6= 0}, we need a condition which
2. SPLITTING ON THE LEVEL SETS OF THE HAMILTONIAN
61
is reminiscent of (4.2), in the sense that it regards again the “amount of the critical points of H”. This is precisely the weak Sard property in (4.11). However we notice that condition (4.11) is much weaker than the previous one; moreover, two examples (for which we refer to [5]) indicate that the weak Sard property is necessary in order to obtain uniqueness. We present here a detailed account of our proof in the basic case of a bounded divergence-free vector field: we remark that various generalizations and extensions are possible (we refer again to [5]).
2. Splitting on the level sets of the Hamiltonian We will be concerned with the Cauchy problem for the transport equation ( ∂t u + b · ∇u = 0 u(0, ·) = u ¯
(4.4)
in D0 ([0, T ] × R2 ),
where u ¯ ∈ L∞ (R2 ). We assume that b ∈ L∞ (R2 ; R2 ) has compact support and that div b = 0. Recall that this implies the existence of a Hamiltonian function H ∈ Lipc (R2 ) which satisfies (4.1); we denote by Lipc (X) the space of Lipschitz functions with compact support defined on X. We remark that the compactness of the support has been assumed in order to simplify the classification of the level sets in Theorem 4.5 and to avoid localizations in the equation, but it is not really essential for our problem. We consider the weak formulation of (4.4): u(t, x) ∈ L∞ ([0, T ] × R2 ) is a weak solution of (4.4) if for every ϕ(t, x) ∈ Lipc ([0, T [×R2 ) we have Z
T
Z
Z
u ∂t ϕ + b · ∇ϕ dxdt = −
(4.5) R2
0
u ¯ϕ(0, ·) dx . R2
We now recall a particular case of the coarea formula (A.3). For every h ∈ R we denote by Eh the level sets Eh = x ∈ R2 : H(x) = h . Then, for H ∈ Lipc (R2 ) and for every function φ ∈ L1 (R2 ), we have Z Z
Z (4.6)
φ dx = R2 ∩{∇H6=0}
R
Eh
φ 1 dH dh . |∇H|
62
4. WELL-POSEDNESS IN THE TWO-DIMENSIONAL CASE
Using (4.6) in (4.5) and recalling that b = ∇⊥ H we obtain Z Z TZ u ¯ϕ(0, ·) dx 0= u∂t ϕ dxdt + {∇H=0}
{∇H=0} T Z
0
Z
T
{∇H6=0}
{∇H6=0}
0
Z
u ¯ϕ(0, ·) dx
u ∂t ϕ + b · ∇ϕ dxdt +
+ (4.7)
Z
Z
Z
u ¯ϕ(0, ·) dx
u∂t ϕ dxdt +
= 0
Z
{∇H=0}
{∇H=0} T Z Z
+ 0
R
Z Z + R
Eh
u 1 ∂t ϕ + b · ∇ϕ dH dhdt Eh |∇H| u ¯ ϕ(0, ·) dH 1 dh . |∇H|
The following lemma will allow the selection of the level sets of the Hamiltonian. We recall that we denote by H# L 2 the push-forward of the Lebesgue measure on R2 via the function H, defined according to (A.1). Lemma 4.2. If u(t, x) is a weak solution of (4.4) and η(h) ∈ L1 (R, H# L 2 ), then u(t, x)η(H(x)) is a weak solution of (4.4). Proof. Considering the weak formulation (4.5) with test function ψ(t, x) = ϕ(t, x)η(H(x)) we deduce the validity of the lemma for any Lipschitz function η. The thesis for every η ∈ L1 (R, H# L 2 ) follows from an approximation procedure, since no derivatives of η are involved in the weak formulation. We now introduce some notation that will be used in the rest of this chapter. We consider the measure λϕ defined by Z T 2 λ ϕ = H# u∂t ϕ dt + u ¯ϕ(0, ·) L {∇H = 0} . 0
It is readily checked that λϕ H# L 2 {∇H = 0} . We denote by λϕ (h) the density 2 of λϕ with respect to H# L {∇H = 0} , i.e. λϕ = λϕ (h)H# L 2 {∇H = 0} . Moreover, for every η ∈ L1 (R, H# L 2 ), we have λη(H)ϕ = η(h)λϕ (h)H# L 2 {∇H = 0} . We perform the decomposition of H# L 2 {∇H = 0} into the absolutely continuous and the singular (with respect to L 1 ) parts: a s H# L 2 {∇H = 0} = H# L 2 {∇H = 0} (h)L 1 + H# L 2 {∇H = 0} .
3. THE WEAK SARD PROPERTY
63
Going back to (4.7), using Lemma 4.2 and the notation introduced we obtain, for every η ∈ L1 (R, H# L 2 ), Z η(h)λϕ (h)d H# L 2 {∇H = 0} (h) R Z Z TZ u 1 + η(h) dhdt ∂t ϕ + b · ∇ϕ dH (4.8) 0 R Eh |∇H| Z Z u ¯ η(h) + ϕ(0, ·) dH 1 dh = 0 . R Eh |∇H| The arbitrariness of the function η ∈ L1 (R, H# L 2 ) in (4.8) then gives the following: (i) for L 1 -a.e. h ∈ R we have a λϕ (h) H# L 2 {∇H = 0} (h) Z Z TZ (4.9) u ¯ u ∂t ϕ + b · ∇ϕ dH 1 dt + ϕ(0, ·) dH 1 = 0 ; + |∇H| |∇H| Eh 0 Eh s 2 (ii) for H# L {∇H = 0} -a.e. h ∈ R we have (4.10)
λϕ (h) = 0 . 3. The weak Sard property
We see from equation (4.9) that the dynamics in {∇H 6= 0} and in {∇H = 0} could be coupled: this can actually happen, as shown in the examples constructed in [5]. This means that we can have interactions between the areas in which the velocity is zero and the ones in which it is nonzero. In order to separate the two dynamics we need the following weak Sard property of the Hamiltonian. Definition 4.3. We say that H ∈ Lipc (R2 ) satisfies the weak Sard property if (4.11) H# L 2 {∇H = 0} ⊥ L 1 . Using the notation introduced in the previous section this means that a H# L 2 {∇H = 0} (h) = 0 for L 1 -a.e. h ∈ R. The connection with the classical Sard theorem (see for instance Theorem 3.4.3 of [91]) is evident: here we are requiring that the “image” (via the push-forward through H) of the Lebesgue measure L 2 restricted to the set of the critical points {∇H = 0} is “not seen” by the Lebesgue measure L 1 in the codomain. Assuming the weak Sard property we can separate the two dynamics, hence from equations (4.9) and (4.10) we deduce the following result. Theorem 4.4. Let b ∈ L∞ (R2 ; R2 ) with compact support and assume that div b = 0. Let H ∈ Lipc (R2 ) be as in (4.1) and assume that H satisfies the weak Sard property (4.11). Let u ∈ L∞ ([0, T ] × Rd ) be a weak solution of (4.4). Then we have Z TZ Z u u ¯ (4.12) ∂t ϕ + b · ∇ϕ dH 1 dt + ϕ(0, ·) dH 1 = 0 |∇H| |∇H| 0 Eh Eh
64
4. WELL-POSEDNESS IN THE TWO-DIMENSIONAL CASE
for L 1 -a.e. h ∈ R and Z TZ (4.13) 0
Z u ¯ϕ(0, ·) dx = 0 .
u∂t ϕ dxdt + {∇H=0}
{∇H=0}
Notice that (4.13) gives u(t, x) = u ¯ for L 1 ⊗ L 2 -a.e. (t, x) ∈ [0, T ] × {∇H = 0}. This means that, thanks to the weak Sard property, the uniqueness for the Cauchy problem (4.4) is equivalent to the uniqueness for the “reduced problems” (4.12) on the level sets, for L 1 -a.e. h ∈ R. The issue of the uniqueness on the level sets is discussed in Section 4.5. We stress the fact that the weak Sard property (4.11) is not just a technical assumption that we need in order to implement our strategy: it is really necessary for the well-posedness, as pointed out by the two counterexamples of [5], where nonuniqueness phenomena are shown in the case of vector fields which do not satisfy (4.11). 4. Structure of the level sets In this section we give a detailed description of the structure of the level sets Eh = {x ∈ R2 : H(x) = h} . We first notice, by the continuity of H and by the assumption of compactness of the support, that for every h 6= 0 the set Eh is compact. Moreover, applying the coarea formula (4.6) with φ ≡ 1 we obtain Z Z 1 1 dH dh = L 2 {∇H 6= 0} < +∞ , R Eh |∇H| by the compactness of spt H. This implies that Z 1 (4.14) dH 1 < +∞ for L 1 -a.e. h ∈ R. |∇H| Eh In particular, since |∇H| ≤ kbk∞ , this yields H 1 (Eh ) < +∞
for L 1 -a.e. h ∈ R.
For every h ∈ R, we denote by Ch the family of all the connected components C of Eh such that H 1 (C) > 0 (in fact, these are just the connected components which contain more than one point). We collect together in the following theorem all the results relative to the classification of the level sets. For the proof we refer to [5]. See Appendix 1 for the notion of rectifiable set. Theorem 4.5. Let H ∈ Lipc (R2 ). For L 1 -a.e. h ∈ R the following statements hold. (i) Eh is H 1 -rectifiable and H 1 (Eh ) < +∞; the map H is differentiable in x and ∇H 6= 0 at H 1 -a.e. x ∈ Eh ; the function 1/|∇H| belongs to L1 (Eh , H 1 ). (ii) The family Ch is countable and H 1 (Eh \ ∪C∈Ch C) = 0.
5. UNIQUENESS ON THE LEVEL SETS
65
(iii) Every C ∈ Ch is a closed simple curve. More precisely, it is possible to find a Lipschitz injective parametrization γ : [α, β]∗ → C such that (4.15)
γ(s) ˙ = ∇⊥ (γ(s))
for L 1 -a.e. s ∈ [α, β]∗ ,
where we denote by [α, β]∗ the quotient space consisting of the interval [α, β] with identified endpoints, endowed with the distance dist [α,β]∗ (x, y) = min |x − y|, (β − α) − |x − y| . We will also need the following topological lemma. Lemma 4.6. Let h ∈ R such that the conclusions of Theorem 4.5 hold. Then, for every C ∈ Ch , there exists a decreasing sequence {Un } of bounded open sets in R2 such that ∂Un ∩ Eh = ∅ for every n and Eh ∩ (∩n Un ) = C. 5. Uniqueness on the level sets We first show, using Lemma 4.6, that equation (4.12) can be separated into a family of equations on the connected components of each level set. Proposition 4.7. Let b ∈ L∞ (R2 ; R2 ) with compact support and assume that div b = 0. Let H ∈ Lipc (R2 ) be as in (4.1) and assume that H satisfies the weak Sard property (4.11). Let u ∈ L∞ ([0, T ] × Rd ) be a weak solution of (4.4). Then, for every C ∈ Ch , for L 1 -a.e. h ∈ R, we have Z Z TZ u ¯ u ∂t ϕ + b · ∇ϕ dH 1 dt + ϕ(0, ·) dH 1 = 0 . (4.16) |∇H| |∇H| C 0 C Proof. We fix h ∈ R such that the conclusions of Theorems 4.4 and 4.5 hold. We choose a sequence {Un } as in Lemma 4.6. Since ∂Un and Eh are compact sets we have dist (∂Un , Eh ) = n > 0 . Thus we fix a standard convolution kernel ρ with spt ρ ⊂ B1 (0) and for every n we set γn (x) = 1Un ∗ ρn /4 (x) . We rewrite equation (4.12) with the test function ϕ(t, x)γn (x). We have Z Z TZ u ¯ u 0= ∂t ϕγn + b · ∇ ϕγn dH 1 dt + ϕ(0, ·)γn dH 1 |∇H| |∇H| Eh 0 Eh Z Z TZ u u ¯ = ∂t ϕ + b · ∇ϕ dH 1 dt + ϕ(0, ·) dH 1 . |∇H| |∇H| Eh ∩Un 0 Eh ∩Un We now let n → ∞ in the above equality. Recalling (4.14) and applying the Lebesgue dominated convergence theorem we eventually obtain (4.16). We are now in the position to formulate and prove our main result. Theorem 4.8. Let b ∈ L∞ (R2 ; R2 ) with compact support and assume that div b = 0. Let H ∈ Lipc (R2 ) be as in (4.1) and assume that H satisfies the weak Sard property (4.11). Then, for every initial data u ¯ ∈ L∞ (R2 ), the Cauchy problem (4.4) has a unique ∞ 2 solution u ∈ L ([0, T ] × R ).
66
4. WELL-POSEDNESS IN THE TWO-DIMENSIONAL CASE
Proof. The existence is obtained by the same regularization technique used in Theorem 2.3. To show the uniqueness, by linearity it suffices to show that the only solution u ∈ L∞ ([0, T ] × R2 ) with initial data u ¯ ≡ 0 is u ≡ 0. Recalling the discussion at the end of Section 4.3 and the result of Proposition 4.7, it is enough to show that, for every C ∈ Ch , for L 1 -a.e. h ∈ R, the validity of Z TZ u ∂t ϕ + b · ∇ϕ dH 1 dt = 0 0 C |∇H| for every ϕ(t, x) ∈ Lipc ([0, T [×R2 ) implies u(t, x) = 0 for L 1 ⊗ H 1 -a.e. (t, x) ∈ [0, T ] × C. We proceed in several steps. Step 1. Parametrization of C. We fix h ∈ R such that the conclusions of Theorem 4.5 and of Proposition 4.7 hold. Using the result in Theorem 4.5 (iii) we know that every C ∈ Ch is a closed simple curve and we choose a Lipschitz injective parametrization γ : [α, β]∗ → C which satisfies (4.15). Hence from (4.16) (with u ¯ ≡ 0) we get Z TZ β u(t, γ(s)) (∂t ϕ)(t, γ(s)) + b(γ(s)) · (∇ϕ)(t, γ(s)) dsdt = 0 (4.17) 0
α
for every ϕ ∈ Lipc ([0, T [×R2 ). We set (4.18)
ϕ(t, ˜ s) = ϕ(t, γ(s)) .
Differentiating both sides of (4.18) with respect to s we get ∂s ϕ(t, ˜ s) = (∇ϕ)(t, γ(s)) · γ(s) ˙ = (∇ϕ)(t, γ(s)) · (∇⊥ H)(γ(s)) , and this implies (4.19)
b(γ(s)) · (∇ϕ)(t, γ(s)) = (∇⊥ H)(γ(s)) · (∇ϕ)(t, γ(s)) = ∂s ϕ(t, ˜ s) .
Setting u ˜(t, s) = u(t, γ(s)) and inserting (4.19) in (4.17) we obtain Z TZ β (4.20) u ˜ ∂t ϕ˜ + ∂s ϕ˜ dsdt = 0 0
α
for every ϕ˜ : [0, T ] × [α, β]∗ → R of the form ϕ(t, ˜ s) = ϕ(t, γ(s)) for some ϕ ∈ Lipc ([0, T [×R2 ). Step 2. Test functions in [0, T ] × [α, β]∗ . We notice that, up to now, we cannot see (4.20) as a distributional equation on [0, T ] × [α, β]∗ : indeed, we are allowed to use as test functions only the particular ϕ’s ˜ of the form above. However, the following lemma from [5] holds. Lemma 4.9. Every ψ ∈ Lipc ([0, T [×, [α, β]∗ ) can be approximated uniformly with a sequence of functions {ϕ˜n } of the form above and such that Lip(ϕ˜n ) is equi-bounded. This means that we can write (4.20) with ϕ˜ = ϕ˜n for every n and passing to the limit we get Z TZ β (4.21) u ˜ ∂t ψ + ∂s ψ dsdt = 0 0
α
for any ψ ∈ Lipc ([0, T [×, [α, β]∗ ). This is now a distributional equation on [0, T ]×[α, β]∗ .
5. UNIQUENESS ON THE LEVEL SETS
67
Step 3. Uniqueness on C. Now it suffices to notice that (4.21) is the distributional form of the Cauchy problem ( ∂t u ˜ + ∂s u ˜=0 (4.22) u ˜(0, ·) = 0 . By the smooth theory for the transport equation of Section 1.4 we know that the only solution to this problem is u ˜ ≡ 0. From the definition of u ˜ we see that this means that u(t, x) = 0 for L 1 ⊗ H 1 -a.e. (t, x) ∈ [0, T ] × C. We conclude the desired thesis. We close this chapter by presenting a particular case in which the weak Sard property (4.11) is satisfied by the function H ∈ Lipc (R2 ) associated to b as in (4.1) and thus the uniqueness result of Theorem 4.8 holds. See [5] for the proof. Corollary 4.10. Let b ∈ L∞ (R2 ; R2 ) with compact support and assume that div b = 0 and that b is approximately differentiable L 2 -a.e. in R2 . Then, for every initial data u ¯ ∈ L∞ (R2 ), the Cauchy problem (4.4) has a unique solution u ∈ ∞ 2 L ([0, T ] × R ). We observe that the approximate differentiability assumption on b in Corollary 4.10 is of “qualitative” type, in contrast with the weak regularity assumptions (for instance Sobolev or BV ) in Chapter 2.
CHAPTER 5
Counterexamples to the well-posedness and related constructions In this chapter we present some counterexamples, whose aim is to show the sharpness of the assumptions made in the various renormalization theorems. In the first section we illustrate an elegant construction by Depauw [81], which yields an explicit example of nonuniqueness for a vector field enjoying a regularity very close to the one needed in Ambrosio’s theorem. This example has striking consequences on the possibility of building a theory for multi-dimensional systems of conservation laws based on transport equations: this is obtained in a joint work with De Lellis [65] and is briefly presented in Section 5.2. In Section 5.3 we make some comments on two counterexamples to the uniqueness (in fact at the ODE level) by DiPerna and Lions [83]. The last section is devoted to some examples by Colombini, Luo and Rauch [62], relative to the lack of propagation of classical regularities for vector fields which are less than Lipschitz.
1. Depauw’s counterexample Following some ideas of Aizenman ([2]), Depauw in [81] and Colombini, Luo and Rauch in [61] have recently given some counterexamples to the uniqueness for the transport equation with a bounded divergence-free vector field (see also [48] for a related construction). These examples also show the sharpness of the result by Ambrosio in Theorem 2.12: for instance, the vector field constructed in [81] is “almost BV ”, in a sense that will be clear in a moment. The example of Depauw consists of a bounded planar divergence-free vector field a(t, x) with two different bounded distributional solutions of ( ∂t w + div (aw) = 0 (5.1) w(0, ·) = 0 . We want to show briefly the construction of this vector field. First of all we define b : [−1/2, 1/2]2 → R2 as if 0 < |x2 | < |x1 | < 1/4 (0, 4x1 ) (−4x2 , 0) if 0 < |x1 | < |x2 | < 1/4 (5.2) b(x1 , x2 ) = 0 otherwise 69
70
5. COUNTEREXAMPLES TO THE WELL-POSEDNESS AND RELATED CONSTRUCTIONS
1/2
1/2
Figure 1. Depauw’s vector field. and we extend it periodically to R2 (see Figure 1). The field a(t, x) is then given by 0 if t < 0 or t > 1 a(t, x) = j b(2 x) if t ∈ Ij = 2−j ( 21 , 1) for some j ∈ N , and we also define c(t, x) = a(1 − t, x). It follows immediately that a and c are bounded and divergence-free. Moreover, note that for L 1 -a.e. t ∈ R we have a(t, ·) ∈ BVloc (R2 ; R2 ), but a does not belong to L1 ([0, α]; BVloc (R2 )) for any α > 0. We want to describe the flow of c. First of all, we recall that for 0 ≤ t0 < t1 < 1 we denote by X (c) (t1 , t0 , x) the solution at time t1 of the problem ( ˙ ψ(t) = c(t, ψ(t)) ψ(t0 ) = x . Note that X (c) is well-defined since c is piecewise smooth on [t0 , t1 ]×R2 . Next we let u0 2 be the Z2 -periodic function given by u0 (x1 , x2 ) = sgn (x1 x2 ) on the square − 21 , 12 . If we define uk (x) = u0 (2k x), then uk X (c) 1 − 2−k , 1 − 2−k−1 , x = uk+1 (x) (see Figure 2). By the semigroup property of X(c) we conclude that u0 (X (c) (0, 1−2−k , x)) = uk (x). Therefore u(t, x) = u0 X (c) (0, t, x) is a bounded distributional solution of ∂t u + div (uc) = 0 (5.3) u(0, ·) = u0 . Note that u(t, ·) converges weakly (but not strongly) to 0 as t ↑ 1. Therefore w defined by w(t, x) = u(1 − t, x) is a nontrivial weak solution of (5.1). 2. Oscillatory solutions to transport equations The example of [81] presented in Section 5.1 has severe consequences on the possibility of building a theory for multi-dimensional systems of conservation laws based on
2. OSCILLATORY SOLUTIONS TO TRANSPORT EQUATIONS
71
1/2
1/2
1/2
1/2
Figure 2. The effect of Depauw’s vector field b acting for a time of 1/2 on a chessboard of side 1/4. transport equations. We recall the Keyfitz and Kranzer system d X ∂ ∂t u + f j (|u|)u = 0 u : [0, T ] × Rd → Rk (5.4) ∂xj j=1 u(0, ·) = u ¯ already presented in the Introduction. For every j = 1, . . . , d the map f j : R+ → R is smooth. Notice the very special feature of this system: the nonlinearity only depends on the moduls of the solution. Even in the presence of this special “triangular” structure, Bressan [48] proved that the Cauchy problem (5.4) can be ill posed in L∞ , showing initial data u which generate renormalized entropy solutions with wild oscillations. We can decouple (5.4) in a scalar conservation law for the modulus ρ = |u| (5.5) ∂t ρ + div f (ρ)ρ = 0 and a continuity equation for the angular part θ = u/|u| (5.6) ∂t (ρθ) + div f (ρ)ρθ = 0 . The heuristic idea behind the well-posedness result for (5.4) in [15] and [11] is to consider first Kruˇzkov’s entropy solutions of (5.5). If |¯ u| ∈ BV (Rd ) then from the theory of entropy solutions ([102]) we have ρ(t, ·) ∈ BV (Rd ) for every t ∈ [0, T ]. Then we consider the continuity equation (5.6) for the vector unknown θ: the vector field f (ρ) belongs to BV ([0, T ] × Rd ) and (5.5) means that this vector field is nearly incompressible. We observe that the theory of renormalized solutions for nearly incompressible vector fields presented in Section 2.8 does not cover immediately this case (the renormalization property in this general framework is presently an open problem: see Conjecture 6.18). However, some ad hoc arguments based on additional properties of the vector field f (ρ) (namely the fact that (5.5), as a continuity equation in ρ with f (ρ) fixed, possesses a solution enjoying BV regularity) are enough to show well-posedness for (5.6), in the
72
5. COUNTEREXAMPLES TO THE WELL-POSEDNESS AND RELATED CONSTRUCTIONS
class of the so-called renormalized entropy solutions (see [15], [11] and [76] for the details). It is therefore tempting to try to generalize this approach to more general systems of conservations laws. Another kind of prototype of multi-dimensional system consists in two continuity equations coupled through some nonlinearity: ∂ u + div f (v)u =0 t ∂ v + div g(u)v = 0 f, g : R → Rd t (5.7) u, v : [0, T ] × Rd → R . u(0, ·) = u ¯ v(0, ·) = v¯ The natural attempt would be the implementation of an iterative procedure. We first find u1 and v2 solutions to the two continuity equations with fixed vector fields: ∂t u1 + div f (¯ v )u1 = 0 ∂ v + div g(¯ u)v1 = 0 t 1 u1 (0, ·) = u ¯ v2 (0, ·) = v¯ . We iterate this procedure, finding for every k ∈ N solutions uk and vk to the continuity equations in which the vector fields are fixed and depend on the solutions uk−1 and vk−1 found in the previous step: ∂ u + div f (v )u =0 t k k−1 k ∂ v + div g(u )v = 0 t k k−1 k u (0, ·) = u ¯ k vk (0, ·) = v¯ . Then one might expect convergence of the two sequences {uk } and {vk } to a solution of (5.7). To do this, we need to identify a function space S ⊂ L∞ (a) which embeds compactly in L1loc , in order to be able to show the existence of the limit, at least up to subsequences; (b) which contains BV , because equations of this type, already in the scalar case, develop singularities along surfaces of codimension one (shocks) in finite time; (c) such that every transport equation with initial data in S and with vector field belonging to S and with bounded divergence possesses a solution belonging to S , to be sure that at each passage of the iteration we can choose a solution which belongs to our space. The question of the existence of such a space was originally raised by Bressan. However, in a joint work with De Lellis [65], we are able to show that such a space does not exist. We will not be more precise on the statement of this result, for which we refer to [65]. Instead we prefer to describe the idea of our proof and its connection with the counterexample in Section 5.1. A small modification of the construction by Depauw yields an autonomous divergence-free vector field u ∈ L∞ , an initial data v ∈ L∞ ∩ BV , and a time T > 0 such
3. TWO COUNTEREXAMPLES BY DIPERNA AND LIONS
73
that in [0, T ] × R3 there exists a unique bounded weak solution of ∂t v + div (uv) = 0 (5.8) v(0, ·) = v¯ , and v(t, ·) converges to v(T, ·) weakly∗ in L∞ but not strongly in L1loc . With an explicit construction it is possible to show the existence of vector fields fi ∈ BV ∩ L∞ with div fi ∈ L∞ and initial data w ¯i ∈ BV ∩ L∞ such that if we consider the unique weak solutions of ∂t wi + div (fi wi ) = 0 (5.9) wi (0, ·) = w ¯i , then wi (1, ·) is the i-th component ui of Depauw’s vector field. This means that, if we assume the existence of a space S satisfying our requirements, then necessarily fi , w ¯i ∈ S , because of (b). From the closure property (c) and thanks to the fact that the solutions to the systems (5.9) are unique we deduce that ui = wi (1, ·) ∈ S . Since, again from (b), v¯ ∈ S , we consider (5.8) and deduce from (c) that v ∈ S . But we already noticed that v(t, ·) converges to v(T, ·) weakly∗ in L∞ but not strongly in L1loc , and this violates the compactness condition (a). Hence we deduce that such a space S cannot exist. This result is clearly a counterexample to the strategy (it is not possible to solve multi-dimensional systems via reduction to transport equations), but not to the wellposedness result. It would be very interesting to understand whether one can use similar constructions to produce hyperbolic systems of conservation laws ∂t U + div F (U ) = 0 and BV initial data with highly oscillatory admissible solutions. Slight modifications of our example produce fluxes F such that each DFi is triangular, but the corresponding systems are not hyperbolic. 3. Two counterexamples by DiPerna and Lions In this section we comment two counterexamples to the uniqueness presented by DiPerna and Lions in [83]. These counterexamples are in the spirit of the ODE framework: the “right” notion of solution of the ODE in this relaxed setting will be presented in Section 6.3, but roughly speaking we require that the flow does not concentrate trajectories, in the sense that the push-forward of the Lebesgue measure L d via the flow X(t, ·) is absolutely continuous with respect to L d and has bounded density (this is the so-called regular Lagrangian flow, compare in particular with Definition 6.4). 3.1. A W 1,p vector field with unbounded divergence. The first example of [83] is relative to a vector field with Sobolev regularity but with unbounded divergence, for which an infinite family of solutions to the ODE is constructed. The construction of [83] (taken from a previous work by Beck [30]) is two-dimensional, but since the vector field is constant in the second direction the phenomenon is intrinsically onedimensional, thus we are going to describe it as a flow on the real line. Moreover, it has been pointed out by Ambrosio that all but one among the constructed flows create concentration of trajectories. Thus this example is not a counterexample to the uniqueness of the regular Lagrangian flow. It would be interesting to understand
74
5. COUNTEREXAMPLES TO THE WELL-POSEDNESS AND RELATED CONSTRUCTIONS
whether it is possible to have examples of nonuniqueness of the solution even in the cases where the uniqueness of the regular Lagrangian flow is already known. For any p ∈ [1, +∞[ fixed, we construct a uniformly continuous vector field b ∈ 1,p Wloc ∩ L∞ (R; R) and infinitely many solutions to the ODE with vector field b. Let us ∞ consider a compact set K ⊂ [0, 1] with empty interior and R x a function g ∈ C (R) with 0 ≤ g < 1 and such that {g = 0} = K. We set f (x) = 0 g(s) ds. Assuming that g(s) converges to 1 for |s| → ∞ we obtain that f is bijective. We also consider A = {µ ∈ M+ (K) : µ does not possess atoms} . For every µ ∈ A we set Λ−1 µ (x) = x + µ([0, x]) ,
fµ = f ◦ Λ µ .
Noticing that Λ−1 µ is strictly increasing and continuous we deduce that fµ is well-defined, continuous and strictly increasing. Moreover, Λµ is 1-Lipschitz. We also set b(x) = f 0 (f −1 (x)) ,
Xµ (t, x) = fµ (t + fµ−1 (x)) .
Since f and fµ are strictly increasing and continuous, we obtain that b and Xµ are well-defined and continuous. Moreover it is immediate to check that Xµ satisfies the semigroup property. We want to show that Xµ is everywhere differentiable with respect to t (equivalently, that fµ is everywhere differentiable) and that (5.10)
∂Xµ (t, x) = fµ0 (t + fµ−1 (x)) = fµ0 (fµ−1 (Xµ (t, x))) ∂t (∗)
= f 0 (f −1 (Xµ (t, x))) = b(Xµ (t, x)) .
Notice that, once we know that fµ is differentiable, (∗) is the only equality in (5.10) which is not an immediate consequence of the definitions given above. The validity of (5.10) will imply that, for every µ ∈ A, the map Xµ is a flow relative to the vector field b. We claim that fµ is differentiable and that for every z ∈ R the following equality holds: (5.11)
fµ0 (z) = f 0 (Λµ (z)) .
Using (5.11) with z = fµ−1 (Xµ (t, x)) and recalling that Λµ ◦ fµ−1 = f −1 we see that (5.11) implies the validity of the equality marked by (∗) in (5.10). We now prove (5.11) considering two cases. Case 1. If Λµ (z) 6∈ K, then for s sufficiently close to z we have Λ−1 µ (Λµ (z) + s − z) = s , since the complement of K is an open set. Hence fµ is differentiable in z with fµ0 (z) = f 0 (Λµ (z)), and (5.11) holds for such a z. Case 2. If Λµ (z) ∈ K, recalling that Λµ is 1-Lipschitz, for s is sufficiently close to z we obtain |fµ (s) − fµ (z)| = |f (Λµ (s)) − f (Λµ (z))| ≤ C|Λµ (s) − Λµ (z)|2 ≤ C|s − z|2 ,
3. TWO COUNTEREXAMPLES BY DIPERNA AND LIONS
75
since f 0 (Λµ (z)) = 0. Hence fµ0 (z) = 0 = f 0 (Λµ (z)) and we eventually get (5.11) also in this case. 1,p We now show how it is possible to choose g in such a way that b ∈ Wloc (R; R) for an arbitrary p ∈ [1, +∞[. Noticing that b0 (x) = g 0 (f −1 (x))[g(f −1 (x))]−1 we can compute Z Z Z |b0 (x)|p dx = |g 0 (f −1 (x))|p |g(f −1 (x))|−p dx = |g 0 (y)|p |g(y)|−(p−1) dy . R
R
R
Therefore, it is enough to choose g0 satisfying all the previous requirements imposed on g and eventually set g = g0p . Since this implies |g|−(p−1) |g 0 |p = pp |g00 |p , it is enough to choose g0 such that g00 ∈ Lp (R). Notice also that div b(x) = b0 (x) does not belong to L∞ (R). We finally show why the only flow Xµ which does not create concentration of trajectories is the one corresponding to µ = 0. We preliminarily show that (i) The set f (K) has zero Lebesgue measure; (ii) The set Λ−1 µ (K) has strictly positive Lebesgue measure, if µ 6= 0. We observe that the statement in (i) is immediate, since f is smooth and K consists of the critical points of f . To show (ii) we notice that Λ−1 µ ([0, 1]) = [0, 1 + m] ,
(5.12)
where m = µ([0, 1]). But since the derivative of Λ−1 µ is less or equal to 1 on [0, 1] \ K we also have 1 (5.13) L 1 Λ−1 µ ([0, 1] \ K) ≤ L [0, 1] \ K ≤ 1 . From (5.12) and (5.13) we deduce L 1 Λ−1 µ (K) ≥ m ,
(5.14)
and this proves (ii). Using (5.14) and the equivalence Xµ (t, x) ∈ f (K)
−1 t ∈ Λ−1 µ (K) − fµ (x)
⇐⇒
we eventually obtain the equality Z Z Z −1 (5.15) 1f (K) (Xµ (t, x)) dtdx = L 1 Λ−1 µ (K) − fµ (x) dx = +∞ R
R
R
for every µ ∈ A with µ 6= 0. This implies the existence of two sets A, B ⊂ R with strictly positive Lebesgue measure such that Xµ (t, x) ∈ f (K)
for every (t, x) ∈ A × B,
thus we have concentration of trajectories in the negligible set f (K) for a set of times with positive measure. Moreover, it is possible to check that the density of the measure R \ f (K) Xµ (t, ·)# L 1 does not belong to L∞ loc (R).
76
5. COUNTEREXAMPLES TO THE WELL-POSEDNESS AND RELATED CONSTRUCTIONS
This example should be compared with the much simpler one presented in Remark 6.6, relative to the square root case. One difference is that in the case presented here each flow Xµpsatisfies the semigroup property, while every flow relative to the vector field b(x) = |x| which “stops” in the origin for a positive time does not satisfy the semigroup property. 3.2. A divergence-free vector field with derivatives of fractional order. The second example of [83] regards an autonomous divergence-free vector field b in R2 s,1 such that b ∈ Wloc (R2 ; R2 ) for every s ∈ [0, 1[ and b ∈ Lp (R2 ; R2 )+L∞ (R2 ; R2 ) for every p ∈ [1, 2[. For the definition and the main properties of Sobolev spaces with fractional order we refer to [1] and [123]. We will construct two regular Lagrangian flows (which in fact preserve the Lebesgue measure) relative to this vector field. One might wonder if this lack of derivative which leads to nonuniqueness could be compensated by an enlargement of the summability exponent, i.e. if we have uniqueness for vector fields in W s,p with s < 1 and p = p(s) sufficiently large. This is not at all the case: some variations of the counterexamples of [5] show that we cannot hope for uniqueness even for vector fields belonging to all the H¨older spaces C α (R2 ; R2 ) with α ∈ [0, 1[. Since we want to construct a divergence-free vector field we start by defining the following Hamiltonian function H : R2 → R. We set if |x| < |y| −x/|y| −(x − |y| + 1) if x > |y| H(x, y) = −(x + |y| − 1) if x < −|y| . Thus we define b as ! x ∂H b1 (x, y) = − ∂y = −sgn y |y|2 1{|x|≤|y|} + sgn x 1{|x|>|y|} ! ∂H 1 b2 (x, y) = ∂x = − |y| 1{|x|≤|y|} + 1{|x|>|y|} . The only non trivial verification is the W s,1 regularity of b: we postpone it to the ˜ relative end of this subsection. We now explicitly define two different flows X and X to b. By simmetry it suffices to define the flows for initial data in the set Q = {(x, y) ∈ R2 : x > 0 ,
y > 0,
When x > y we set ( x−t ˜ X1 (t, x, y) = X1 (t, x, y) = x − 2y + t ˜ 2 (t, x, y) = y − t . X2 (t, x, y) = X When x < y we set X1 (t, x, y) = x |y 2 − 2t|1/2 y X (t, x, y) = σ|y 2 − 2t|1/2 2
x 6= y} .
if t ≤ y if t ≥ y
3. TWO COUNTEREXAMPLES BY DIPERNA AND LIONS
77
and X ˜ 1 (t, x, y) = σ x |y 2 − 2t|1/2 y X ˜ (t, x, y) = σ|y 2 − 2t|1/2 , 2
where σ = 1 if t ≤ y 2 /2 and σ = −1 if t ≥ y 2 /2. It is immediate to check that X and ˜ are two different regular Lagrangian flows relative to b and that they preserve the X Lebesgue measure in R2 . s,1 Let us finally check that b ∈ Wloc (R2 ; R2 ). We will give only a sketch of the proof, which uses some technical results from the theory of Besov spaces: see Section 11 of [123], [79] and [124] for the precise statements of the results required here. First of all we observe that s B1,1 (R2 ) ⊂ W s,1 (R2 ) , s (R2 ) is defined according to Definition 10.3 of [123]. Thus where the Besov space B1,1 s around the origin. The following equivalence of norms it suffices to show that b ∈ B1,1 holds:
(5.16)
s kf kB1,1 ∼ kf kB s−1 + kDf kB s−1 . 1,1
1,1
For s < 1, the first term in the right hand side of (5.16) can be dropped. Moreover, when α < 0, we have the equivalence of norms Z 1 d α ∼ (5.17) kgkB1,1 −α kg ∗ ϕ kL1 , 0 where {ϕ } is a standard convolution kernel. Combining all the previous remarks, we need to show that Z 1 d < +∞ . (5.18) 1−s kDb ∗ ϕ kL1 0 We estimate the L1 norm inside the integral as follows: (5.19)
kDb ∗ ϕ kL1 (R2 ) = kDb ∗ ϕ kL1 (B2 (0)) + kDb ∗ ϕ kL1 (R2 \B2 (0)) C ≤ kbkL1 (B3 (0)) + CkDbkM(R2 \B (0)) .
Inserting (5.19) in (5.18) and noticing that kbkL1 (B3 (0)) ∼
and
kDbkM(R2 \B (0)) ∼ log
we deduce that the integral in (5.18) is bounded by Z 1 log 1 C + s d , s 0 which is finite since s < 1. This completes the proof.
78
5. COUNTEREXAMPLES TO THE WELL-POSEDNESS AND RELATED CONSTRUCTIONS
4. Lack of propagation of regularity In the case b ∈ Lip(Rd ; Rd ) we have propagation of the H¨older regularity of the initial data, in the following sense: if u ¯ ∈ C α ∩ L∞ (Rd ) for some α ∈ [0, 1[ and ∞ d u ∈ L ([0, T ] × R ) is the unique solution of the Cauchy problem ( ∂t + b · ∇u = 0 u(0, ·) = u ¯, then u(t, ·) ∈ C α (Rd ) for every t ∈ [0, T ]. Indeed, the solution can be expressed as u(t, x) = u ¯ X(t, ·)−1 (x) and the inverse of the flow X(t, ·)−1 : Rd → Rd is a Lipschitz map (Corollary 1.11). One might wonder if similar results are true for less regular vector fields. In fact, for “classical” regularities, nothing of this sort is true, as pointed out by Colombini, Luo and Rauch [62]. However, in Section 7.5 we will show that a mild regularity, namely Lipexpp regularity, is propagated by W 1,p vector fields with bounded divergence. In the remaining of this section we illustrate the two main examples of [62]. Example 5.1 (Lack of propagation of continuity). Consider a nonnegative function f ∈ C(R) such that f (s) = 0 for s ≤ 0 and f (s) = s(log s)2 for 0 < s ≤ e−2 /2. Assume that f is nondecreasing and uniformly bounded. Set b(x, y) = − f (y), −f (x) . Then b is a bounded divergence-free vector field, which belongs to C α (R2 ; R2 ) for every 1,p (R2 ; R2 ) for every p ∈ [1, +∞[. Now consider any initial data α ∈ [0, 1[ and to Wloc u ¯ ∈ Cc∞ (R2 ) such that u ¯ = 0 if x ≤ 0 and y ≤ 0, and u ¯ > 0 if x > 0 and y > 0 are small. Then, for every relatively open set Ω ⊂ [0, T ] × R2 such that (0, 0, 0) ∈ Ω, the unique solution u ∈ L∞ ([0, T ] × R2 ) with initial data u ¯ is not continuous in Ω. Example 5.2 (Lack of propagation of BV regularity). Consider a function g ∈ Cc (R) such that g(s) = −s log |s| + s for s close to 0 and set b(x, y) = g(y), 0 . p Set r = x2 + y 2 and consider an initial data u ¯ ∈ L∞ ∩ W 1,1 (R2 ) such that 1 for r small. u ¯(x, y) = cos r(log r)2 Then there exists a unique solution u ∈ L∞ ([0, T ] × R2 ) with initial data u ¯, and u(t, ·) 6∈ BVloc (R2 ) for t small.
Part 2
The Lagrangian viewpoint
CHAPTER 6
The connection between PDE and ODE In this chapter we are going to describe the connection between the Eulerian problem (PDE) and the Lagrangian problem (ODE) out of the smooth setting. In Section 1.4 we presented the classical theory of characteristics, which relates the two problems in the case of a sufficiently smooth vector field. Here we are going to present the theory of regular Lagrangian flows, developed by Ambrosio in [8] (see also [13]). In order to study existence, uniqueness and stability of solutions to the ODE, we consider suitable measures in the space of continuous maps, allowing for superposition of trajectories. Then, in some special situations we are able to show that this superposition actually does not occur, but still this “probabilistic” interpretation is very useful to understand the underlying techniques and to give an intrinsic characterization of the flow. We will establish an abstract connection between the well-posedness in the class of bounded solutions for the continuity equation ( ∂t µ + div (bµ) = 0 (6.1) µ0 = µ ¯ and the well-posedness of a suitable notion of solution (the regular Lagrangian flow) of the ordinary differential equation ( γ(t) ˙ = b(t, γ(t)) (6.2) γ(0) = x . With “abstract” we mean that the spirit of these results will be to deduce the wellposedness for the ODE from the well-posedness for the PDE; hence all the results shown in the previous chapters will imply existence, uniqueness and stability for the regular Lagrangian flow. In Section 6.5 we briefly comment on the notion of flow defined by DiPerna and Lions. Section 6.6 is devoted to the case of nearly incompressible vector fields: we follow an approach, due to De Lellis, based on the notion of density of the regular Lagrangian flow. The last section presents Bressan’s compactness conjecture and some recent advances relative to it. 1. Pointwise uniqueness and measure valued solutions We first present a very general criterion, which relates the pointwise uniqueness for the ordinary differential equation with the uniqueness for positive measure-valued solutions to the continuity equation. Notice that, in order to give a meaning to the product bµ when µ is a measure, we assume b to be defined everywhere in [0, T ] × Rd . 81
82
6. THE CONNECTION BETWEEN PDE AND ODE
Theorem 6.1. Let A ⊂ Rd be a Borel set. Then the following two properties are equivalent: (i) Solutions of the ordinary differential equation (6.2) are unique for every initial point x ∈ A; (ii) Positive measure-valued solutions of the continuity equation (6.1) are unique for every initial data µ ¯ which is a positive measure concentrated on A, i.e. such that µ ¯(Rd \ A) = 0. Proof of (ii) ⇒ (i) in Theorem 6.1. This implication is rather easy. Assume that for some x ∈ A there exist two different solutions γ(t) and γ˜ (t) of the ODE starting from x. Then consider µt = δγ(t) and µ ˜t = δγ˜(t) . We clearly have µ0 = µ ˜0 = δx . It is readily checked that µt and µ ˜t are solutions of the continuity equation, but since they are different we are violating assumption (ii). The converse implication is much more complicated and requires the so-called superposition principle, which roughly speaking says that every positive measure-valued solution of the continuity equation can be obtained as a superposition of solutions obtained via propagation along characteristics. The superposition principle will be presented in the next section, together with the proof of the implication (i) ⇒ (ii) of Theorem 6.1. We remark that the applicability of Theorem 6.1 is in fact very limited. On one hand, pointwise uniqueness for the ordinary differential equation is known only under very strong regularity assumptions on the vector field, namely in the cases we presented in Chapter 1 (for instance Lipschitz regularity, one-sided Lipschitz condition, Osgood condition). On the other hand, uniqueness for the continuity equation is known only for particular classes of solutions, tipically for solutions which are bounded functions. It is reasonable that this kind of “weaker PDE uniqueness” should reflect into a weaker notion of uniqueness for the ODE: this leads to the concept of regular Lagrangian flow, which is presented in Section 6.3. 2. The superposition principle In order to understand better the meaning of the superposition principle we recall formula (1.17). If there is a unique flow X(t, x) associated to the vector field b, the only solution of the continuity equation with initial data µ ¯ ∈ M(Rd ) is the measure µt characterized by Z (6.3) hµt , ϕi = ϕ(X(t, x)) d¯ µ(x) ∀ϕ ∈ Cc (Rd ) . Rd
In the following we use the notation ΓT for the space C([0, T ]; Rd ) of continuous paths in Rd . For every x ∈ Rd let us consider a probability measure ηx ∈ P(ΓT ) concentrated on the trajectories γ ∈ ΓT which are absolutely continuous integral solutions of the ordinary differential equation with γ(0) = x. All the families {ηx }x∈Rd in the following discussions are weakly measurable, i.e. for every function Φ ∈ Cb (ΓT ) the map Z x 7→ hηx , Φi = Φ(γ) dηx (γ) ΓT
2. THE SUPERPOSITION PRINCIPLE
83
is measurable. Definition 6.2 (Superposition solution). The superposition solution induced by the family {ηx }x∈Rd is the family of measures µηt x ∈ M(Rd ), for t ∈ [0, T ], defined as follows: Z Z ηx (6.4) hµt , ϕi = ϕ(γ(t)) dηx (γ) d¯ µ(x) ∀ϕ ∈ Cc (Rd ) . Rd
ΓT
Using this notation we can give an alternative interpretation of (6.3). If for every x ∈ Rd the solution of the ODE starting from x is unique, then the only admissible measure ηx in (6.4) is ηx = δX(·,x) . But then we have Z Z Z δX(·,x) hµt ϕ(X(t, x)) d¯ µ(x) , , ϕi = ϕ(γ(t)) dδX(·,x) (γ) d¯ µ(x) = Rd
Rd
ΓT
so in this case we reduce to the “deterministic” formula (6.3). We can regard the superposition solution of Definition 6.2 as a “probabilistic” version of (6.3): if there is more than one solution to the ordinary differential equation, then we define our “averaged push-forward” by substituting the quantity ϕ(X(t, x)) with the average R ΓT ϕ(γ(t)) dηx (γ). Let us check that (6.4) defines a solution of the continuity equation. Since the measure ηx is concentrated on solutions of the ODE starting from x ∈ Rd for L d -a.e. x ∈ Rd , using Fubini’s theorem we deduce that for L d ⊗ ηx -a.e. (x, γ) ∈ Rd × ΓT
γ(t) ˙ = b(t, γ(t))
for L 1 -a.e. t ∈ [0, T ]. We notice that, for any ϕ ∈ Cc∞ (Rd ), the map t 7→ hµηt x , ϕi is Lipschitz. Indeed, we can estimate ϕ(γ(t)) − ϕ(γ(s)) ≤ k∇ϕk∞ kbk∞ |t − s| and this implies Z ηx hµ , ϕi − hµηs x , ϕi = t
ϕ(γ(t)) − ϕ(γ(s)) dηx (γ) d¯ µ(x) Rd ΓT ≤ k∇ϕk∞ kbk∞ µ ¯ spt ϕ + BT kbk∞ (0) |t − s| . Z
Thus the distributional derivative of the map t 7→ hµηt x , ϕi coincides with the pointwise one. But since an immediate computation of the pointwise derivative gives Z Z d d ηx hµt , ϕi = ϕ(γ(t)) dηx (γ) d¯ µ(x) dt dt Rd ΓT Z Z = ∇ϕ(γ(t)) · b(t, γ(t)) dηx (γ) dx Rd ΓT Z = b(t, x) · ∇ϕ(x) dµt (x) , Rd
we obtain that the continuity equation holds. The superposition principle says that, for positive solutions, this construction can be reversed: every positive measure-valued solution µt can be realized as a superposition solution µηt x for some {ηx }x∈Rd as above. The result in the form we are going to present
84
6. THE CONNECTION BETWEEN PDE AND ODE
is a slight generalization of the one proved in Section 8.2 of [18]: since we will need to deal with L∞ solutions of the continuity equation we need to consider here measures µt which are only locally finite. We refer to Section 4 of [13] for the case in which µt ∈ P(Rd ) and the vector field satisfies the global summability assumption Z
T
0
Z Rd
|b(t, x)| dµt (x) dt < +∞ . 1 + |x|
See also Maniglia [110] for an extension to the non homogeneous case. Theorem 6.3 (Superposition principle). Fix a bounded vector field b : [0, T ]×Rd → and let µt ∈ M+ (Rd ) be a positive locally finite measure-valued solution of the continuity equation. Then µt is a superposition solution, i.e. there exists a family {ηx }x∈Rd ⊂ P(ΓT ), with ηx concentrated on absolutely continuous integral solutions of the ODE starting from x, for µ ¯-a.e. x ∈ Rd , such that µt = µηt x for any t ∈ [0, T ]. Rd
Proof. Step 1. Uniform control of the local masses. The measure µt is just locally finite; however, due to the finite speed of propagation of the transport equation (we are assuming that b is uniformly bounded), it is easy to show the existence of a function mR , independent of the time, such that (6.5)
µt (BR (0)) ≤ mR
for every t ∈ [0, T ].
This can be proved by integration on suitable cones, see for instance Lemma 2.11 of [11] (the argument used for bounded solutions remains unchanged in the case of measure-valued solutions). Step 2. Construction of an adapted convolution kernel. R We want to contruct a positive convolution kernel ρ ∈ C k (Rd ) with spt ρ = Rd and Rd ρ(x) dx = 1 in such a way that, for some function m ˜ R , we have (6.6)
µt (BR (0)) ≤ m ˜R
for every t ∈ [0, T ] and every ∈]0, 1[,
where µt = µt ∗ ρ . This can be done as follows. For every R j ∈ N consider a standard convolution kernel ρj with spt ρj = Bj (0), ρj ≥ 0 and Rd ρj (x) dx = 1. We define ρ(x) =
∞ X
cj ρj (x) .
j=1
We want to choose the numbers cj in order to obtain a convolution P kernel ρ which satisfies the desired properties. If we choose cj > 0 for every j and j cj = 1 we obtain that ρ is strictly positive on the whole Rd P and has unit total mass. Moreover, in order to obtain a C k function, we also require j cj kρj kC k < ∞, which is clearly satisfied for every k ∈ N if we choose a sequence cj which goes to zero sufficiently fast.
2. THE SUPERPOSITION PRINCIPLE
85
We now show how it is possible to choose the sequence cj in such a way that the uniform control (6.6) is satisfied. We compute Z Z Z µt (BR (0)) = µt (x) dx = ρ (x − y) dµt (y) dx Z
Z
=
Rd
BR (0) ∞ X
=
j=1 ∞ X
≤
BR (0) ∞ X
Z
cj ρj (x − y) dµt (y) dx =
j=1
Z
ρj (x
cj Rd
− y) dx dµt (y) =
BR (0)
dµt (y) ≤
cj
∞ X j=1 ∞ X j=1
Z
j=1
Rd
BR (0)
BR+j (0)
∞ X
Z
Z
ρj (x − y) dµt (y) dx
cj BR (0)
Rd
Z
Z
cj BR+j (0)
ρj (x − y) dx dµt (y)
BR (0)
cj mR+j .
j=1
This means that we need to choose the sequence cj > 0 in such a way that (6.7)
m ˜R =
∞ X
cj mR+j < ∞
j=1
for every R ∈ N. It can be easily checked that it is enough to choose each cj such that cj mj , cj mj+1 , . . . cj m2j ≤
1 . 2j
Indeed with this choice we can estimate m ˜R =
∞ X j=1
cj mR+j =
R−1 X j=1
cj mR+j +
∞ X j=R
cj mR+j ≤
R−1 X j=1
cj mR+j +
∞ X 1 < ∞. 2j
j=R
c0j
Noticing that, if a sequence cj satisfies (6.7) and 0 < ≤ cj , then also the sequence 0 cj satisfies (6.7), we conclude that it is possible to choose a sequence which satisfies all the requirements we need. Hence we have constructed a kernel ρ with the claimed property (6.6). Step 3. Smoothing. We fix a convolution kernel ρ as in the previous step and we consider b(t, ·)µt ∗ ρ (x) and b (t, x) = . µt (x) = µt ∗ ρ (x) µt (x) It is immediate to check the uniform L∞ bound (6.8) R | b(t, ·)µ ∗ ρ (x)| kbk∞ µt (x) t Rd |ρ (x − y)b(t, y)| dµt (y) |b (t, x)| = ≤ ≤ = kbk∞ µt (x) µt (x) µt (x) and that µt is a solution of the continuity equation with vector field b , that is ∂t µ + div (b µt ) = ∂t µt ∗ ρ + div b(t, ·)µt ∗ ρ = 0 . Since b is globally bounded and C k with respect to the space (it is indeed a ratio of two C k functions and the denominator does not vanish) we can define a unique flow
86
6. THE CONNECTION BETWEEN PDE AND ODE
X (t, x) associated to b , which is globally defined for t ∈ [0, T ]. The representation µt = X (t, ·)# µ0 given by Proposition 1.15 holds and we define ηx = δX (·,x) ∈ P(ΓT ) and η = µ0 ⊗ ηx ∈ M+ (Rd × ΓT ) , where the tensor product by definition satisfies Z Z hη , Φ(x, γ)i = Φ(x, γ) dηx (γ) dµ0 (x) Rd
ΓT
for every function Φ ∈ Cb (Rd × ΓT ) whose support has compact projection over Rd . Step 4. Tightness. We consider the functional Ψ : Rd × ΓT → [0, +∞] defined by Z T 2 Ψ : (x, γ) 7→ |γ(t)| ˙ dt ; 0
the functional takes the value +∞ for all the curves γ ∈ ΓT which do not belong to AC 2 ([0, T ]; Rd ), the space of absolutely continuous maps with square-integrable derivative. For every R > 0 we define the space ΓT,R ⊂ ΓT of the curves starting from BR (0), that is ΓT,R = γ ∈ ΓT : γ(0) ∈ BR (0) , and the localized functional
R
Ψ (x, γ) =
Ψ(x, γ) if (x, γ) ∈ BR (0) × ΓT,R +∞ otherwise.
It is readily checked that ΨR is coercive for every R. For every finite λ > 0 let (x, γ) ∈ {ΨR ≤ λ}. Since λ is finite we deduce that x ∈ BR (0) and that γ(0) ∈ BR (0). Moreover for every s and t ∈ [0, T ] we can compute Z Z t |γ(t) − γ(s)| = γ(τ ˙ ) dτ ≤ s
T
|γ(τ ˙ )|2 dτ
1/2
|t − s|1/2 ≤ λ1/2 |t − s|1/2 .
0
Applying the Ascoli–Arzel`a theorem we deduce that each sublevel is relatively compact in Rd × ΓT , as desired. We define also the truncated measure η ,R = η BR (0) × ΓT,R ∈ M+ (Rd × ΓT ) , which is in fact a finite measure. Now we want to evaluate the value of the truncated measure η ,R on the localized functional ΨR . Using the definition of η , the fact that µt
2. THE SUPERPOSITION PRINCIPLE
87
is a solution of the continuity equation and the uniform bound on b we can estimate Z Z Z ,R R R ,R hη , Ψ i = Ψ (x, γ) dη (x, γ) = Ψ(x, γ) dηx (γ) dµ0 (x) Rd ×ΓT
Z
BR (0)
Z
= BR (0) T Z
Z =
0
0
T
ΓT,R
2 Z TZ ∂X dt dµ0 (x) = (t, x) |b (t, X (t, x))|2 dµ0 (x) dt ∂t 0 BR (0) Z TZ |b (t, x)|2 dµt (x) dt |b (t, x)|2 dµt (x) dt =
X (t,BR (0))
≤ kbk2L∞ ([0,T ]×Rd )
0
Z 0
BR+T kbk∞ (0)
T
dµt BR+T kbk∞ (0) dt ≤ T kbk2L∞ ([0,T ]×Rd ) m ˜ R+T kbk∞ .
Together with the coercivity of ΨR , this gives that, for every fixed R, the family {η ,R } ⊂ M+ (Rd × ΓT ) is tight (see the discussion in Appendix 1). Noticing that this family is also equi-bounded by the simple estimate ˜R, η ,R (Rd × ΓT ) = µ0 (BR (0)) ≤ m we can apply the Prokhorov theorem (Theorem A.1) and deduce that it is relatively sequentially narrowly compact. With a standard diagonal argument (based on the fact that narrow convergence is preserved under the restriction of measures) it is possible to construct a sequence {η i }i ⊂ M+ (Rd × ΓT ) and a measure η ∈ M+ (Rd × ΓT ) such that η i ,R → η R narrowly in M+ (Rd × ΓT ) for every R. For future use, we say that η i converges toward η in the sense of the truncated narrow convergence in M+ (Rd × ΓT ). Step 5. Disintegration and equality µt = µηt x . Up to now we have constructed a measure η ∈ M+ (Rd × ΓT ). It is clear that the marginal on Rd of η is µ ¯. Indeed, for every ∈]0, 1[ we have πRd # η = µ0 and the truncated narrow convergence is inherited by the marginal; this, combined with the fact that µ0 converges to µ ¯ d weakly in R , gives that πRd # η = µ ¯. This implies that we can apply the disintegration theorem recalled in Appendix 1 to obtain η=µ ¯ ⊗ ηx , with ηx ∈ P(ΓT ) for µ ¯-a.e. x ∈ Rd . We notice that in the proof we preferred to show convergence at the level of measures on the product Rd × ΓT rather than convergence of ηx for µ ¯-a.e. x. In this way we get a subsequence {i } along which we have convergence which does not depend on the point x; choosing the limit ηx along a different subsequence for each point x could produce a family {ηx }x∈Rd with non-measurable dependence with respect to x. We now check that µηt x = µt . Indeed, for every ∈]0, 1[, applying the definition of ηx and the fact that µt = X (t, ·)# µ0 , we deduce that Z Z ηx (6.9) hµt , ϕi = ϕ(X (t, x)) dµ0 (x) = ϕ(x) dµt (x) = hµt , ϕi Rd
Rd
88
6. THE CONNECTION BETWEEN PDE AND ODE
for every ϕ ∈ Cc (Rd ). Note that Z Z ϕ(x) dµt (x) → (6.10) Rd
ϕ(x) dµt (x) = hµt , ϕi
Rd µt to
by the truncated narrow convergence of µt . Moreover Z Z ϕ(X (t, x)) dµ0 (x) = ϕ(γ(t)) dη (x, γ) Rd ×ΓT Rd Z Z Z (6.11) → ϕ(γ(t)) dη(x, γ) = ϕ(γ(t)) dηx (γ) d¯ µ(x) = hµηt x , ϕi Rd ×ΓT
Rd η
ΓT
by the truncated narrow convergence of to η along the chosen subsequence. We finally obtain the desired equality µηt x = µt from (6.10) and (6.11). Step 6. η is concentrated on solutions of the ODE. In this last step we want to show that ηx is concentrated on solutions of the ODE for µ ¯-a.e. x ∈ Rd . It is enough to show that Z t Z dη(x, γ) = 0 b(s, γ(s)) ds γ(t) − x − (6.12) BR (0)×ΓT
0
for every t ∈ [0, T ] and for every R > 0. Indeed, for every t ∈ [0, T ] and every R ∈ N this gives a µ ¯-negligible set Nt,R ⊂ BR (0) such that the equality Z t (6.13) γ(t) = x + b(s, γ(s)) ds 0
holds for every x ∈ BR (0) \ Nt,R for ηx -a.e. γ ∈ ΓT . Choosing [ [ N= Nt,R t∈[0,T ]∩Q R∈N
(which is clearly L d -negligible) and using the continuity of each trajectory γ, we deduce that (6.13) holds for every t ∈ [0, T ], for every x ∈ Rd \ N for ηx -a.e. γ ∈ ΓT . Our objective is then to show (6.12) for any t ∈ [0, T ] and any R ∈ N. The technical difficulty is that this test function, due to the lack of regularity of b, is not continuous. We claim that (6.14) Z Z t Z TZ γ(t) − x − dη(x, γ) ≤ a(s, γ(s)) ds |b(s, x) − a(s, x)| dµs (x) ds d BR (0)×ΓT
0
0
R
for every continuous vector field a. We first show how the claimed equation (6.14) implies our thesis. It is enough to choose a sequence {ak } of continuous vector fields such that Z TZ (6.15) |b(s, x) − ak (s, x)| dµs (x) ds → 0 as k → ∞. 0
Rd
This can be done by observing that continuous functions are dense in L1loc ([0, T ]×Rd ), µ) RT with respect to the global norm L1 ([0, T ]×Rd , µ) (here µ = 0 µt dt ∈ M+ ([0, T ]×Rd )). To see this, it is enough to write f ∈ L1loc ([0, T ]×Rd , µ) as a sum of compactly supported functions fi and then mollify each fi with a fixed kernel β , but choosing a parameter
2. THE SUPERPOSITION PRINCIPLE
89
= (i) small enough in such a way that the difference fi − fi ∗ β (i) decays like 2−i and the series of the characteristic functions of spt fi ∗ β (i) is bounded. Assuming the validity of (6.14) we can then estimate Z Z t γ(t) − x − dη(x, γ) b(s, γ(s)) ds BR (0)×ΓT
Z ≤ BR (0)×ΓT Z TZ
≤2
0
Z t Z t γ(t) − x − dη(x, γ) a (s, γ(s)) ds + b(s, γ(s)) − a (s, γ(s)) ds k k 0
0
|b(s, x) − ak (s, x)| dµs (x) ds . Rd
0
In the last inequality we used (6.14) to estimate the first term and the fact that µt = µηt x to compute the second one (which is eventually estimated by the global integral on the whole Rd ). Recalling (6.15) we obtain the desired formula (6.12). It remains to show (6.14). For every ∈]0, 1[ we compute Z Z t γ(t) − x − a(s, γ(s)) ds dη (x, γ) BR (0)×ΓT
Z = BR (0)
Z tZ ≤ 0
0
Z t X (t, x) − x − a(s, X (s, x)) ds dµ0 (x) 0
(b − a)(s, X (s, x)) dµ0 (x) ds
BR (0)
Z tZ ≤ 0
h i (b − a )(s, X (s, x)) + (a − a)(s, X (s, x)) dµ0 (x) ds
BR (0)
Z tZ
h i (b − a )(s, x) + (a − a)(s, x) dµs (x) ds ,
≤ 0
BR+T kbk∞ (0)
where we have used the definition of η , the fact that µt = X (t, ·)# µ0 and we have set a (t, x) = a(t, ·)µt ∗ ρ (x) /µt (x). We estimate the first integral. Setting d = b − a we notice that d(t, ·)µt ∗ ρ , |b − c | = |d | = µt hence Z tZ Z tZ d(s, ·)µs ∗ ρ (x) dx ds |d (s, x)| dµs (x) ds = 0
BR+T kbk∞ (0)
Z tZ
Z
≤ 0
Rd
BR+T kbk∞ (0)
0
BR+T kbk∞ (0)
ρ (x − y) dx|d(s, y)| dµs (y) ds ≤
Z
T
Z |d(s, y)| dµs (y) ds .
0
Rd
Since a is continuous we have that a → a locally uniformly, hence the second integral vanishes as → 0. Passing to the limit in the inequalities above (along the subsequence for which we have truncated narrow converge of η to η) we obtain (6.14). Using the superposition principle we can conclude the proof of Theorem 6.1.
90
6. THE CONNECTION BETWEEN PDE AND ODE
Proof of (i) ⇒ (ii) in Theorem 6.1. Let µt be a positive measure-valued solution of the continuity equation with initial data µ ¯. Applying Theorem 6.3 we deduce that µt = µηt x , with ηx ∈ P(ΓT ) concentrated on the absolutely continuous integral solutions of the ODE starting from x, for every point x ∈ A. But assumption (i) precisely means that, for every x ∈ A, the solution is unique. Hence, for every x ∈ A the measure ηx is a Dirac mass supported on the unique trajectory starting from x, and eventually this gives an explicit formula for the solution µt , which is therefore unique. The importance of the superposition principle also relies in the fact that it will allow, using truncations and restrictions of the measures ηx , several manipulations of solutions of the continuity equation: these constructions are not immediate at the level of the PDE, but they are extremely useful in various occasions, see for instance the proof of Proposition 6.9. This viewpoint is very close in spirit to Young’s theory of generalized surfaces and controls [129], a theory with remarkable applications also to nonlinear PDEs (see [82] and [121]) and to the calculus of variations (see [28]). There are also strong connections with Brenier’s weak solutions of the incompressible Euler equation (see [42]), with Kantorovich’s viewpoint in the theory of optimal transportation (see [86] and [117]) and with Mather’s theory (see [112], [113] and [29]). 3. The regular Lagrangian flow As we remarked at the end of Section 6.1, out of the smooth context of the Cauchy– Lipschitz theory the notion of pointwise uniqueness of the ordinary differential equation is not any more the appropriate one. We need to relax the notion of solution, assuming an additional condition which will ensure uniqueness under the various weak differentiability assumptions presented in the previous chapters. With this new concept of solution we will have uniqueness “in the selection sense”: every time we approximate the vector field b with a smooth sequence bh we obtain that the classical flows Xh associated to them converge to the chosen flow for b (see Remark 6.6 and Theorem 6.12). This notion of solution is encoded in the following definition. Definition 6.4 (Regular Lagrangian flow). Let b : [0, T ] × Rd → Rd be a bounded vector field. We say that a map X : [0, T ] × Rd → Rd is a regular Lagrangian flow relative to the vector field b if (i) for L d -a.e. x ∈ Rd the map t 7→ X(t, x) is an absolutely continuous integral solution of γ(t) ˙ = b(t, γ(t)) for t ∈ [0, T ], with γ(0) = x; (ii) there exists a constant L independent of t such that X(t, ·)# L d ≤ LL d . The constant L in (ii) will be called the compressibility constant of X. Compare this notion of solution with the one by DiPerna and Lions, presented in Section 6.5. We remark that condition (ii) in Definition 6.4 is a uniform compressibility condition: we ask that trajectories do not concentrate, and we quantify this with the compressibility constant L. This condition is also important because we obtain a notion of solution which is invariant under modifications of b on negligible sets: if b(t, x) =
3. THE REGULAR LAGRANGIAN FLOW
91
˜b(t, x) for L d+1 -a.e. (t, x) ∈ [0, T ]×Rd it is easy to check that X is a regular Lagrangian flow relative to b if and only if it is a regular Lagrangian flow relative to ˜b. Remark 6.5 (Lb -Lagrangian flows). It is also possible to define a more general concept of solution, the one of Lb -Lagrangian flow (see Definition 13 of [13]). Consider a convex class Lb of positive measure-valued solutions µt of the continuity equation with vector field b, satisfying the monotonicity assumption 0 ≤ µ0t ≤ µt ∈ Lb
=⇒
µ0t ∈ Lb ,
whenever µ0t is a solution of the continuity equation and satisfies an appropriate integrability condition. Then, given a measure µ ¯ ∈ M+ (Rd ), we can define the notion of Lb -Lagrangian flow starting from µ ¯ requiring that the map X(·, x) is an absolutely continuous integral solution of the ordinary differential equation for µ ¯-a.e. x ∈ Rd and that the measures µt = X(t, ·)# µ ¯ induced via push-forward belong to the class Lb . We can see Lb -Lagrangian flows as suitable selections of solutions of the ordinary differential equation, made in such a way to produce densities in Lb . The notion we have given in Definition 6.4 corresponds to the case Lb = L∞ ([0, T ] × Rd ). We decided here to focus only on this concept of solution in order to simplify the presentation, but also because tipically the well-posedness results at the PDE level can be shown for bounded solution, hence L∞ is the most natural class to be considered (but see also Remark 6.8). Remark 6.6. Let us consider again the square root example (see Example 1.1): ( p γ(t) ˙ = |γ(t)| γ(0) = x0 . Solutions of the ODE are not unique for x0 = −c2 < 0. Indeed, they reach the origin in time 2c, where they can stay for an arbitrary time T , then continuing as γ(t) = 41 (t − T − 2c)2 . Let us consider for instance the Lipschitz approximation (that p could easily be made smooth) of b(γ) = |γ| given by p if −∞ < γ ≤ −2 |γ| b (γ) = if −2 ≤ γ ≤ λ − 2 p γ − λ + 22 if λ − 2 ≤ γ < +∞, with λ − 2 > 0. Then, solutions of the approximating ODEs starting from −c2 reach the value −2 in time t = 2(c − ) and then they continue with constant speed until they reach λ − 2 , in time T = λ /. Then, they continue as λ − 22 + 14 (t − t − T )2 . Choosing λ = T , with T > 0, by this approximation we select the solutions that do not move, when at the origin, exactly for a time p T. Other approximations, as for instance b (γ) = + |γ|, select the solutions that move immediately away from the singularity at γ = 0. Among all possibilities, this family of solutions γ(t, x0 ) is singled out by the property that γ(t, ·)# L 1 is absolutely continuous with respect to L 1 , so that no concentration of trajectories occurs at the origin. To see this fact, notice that we can integrate in time the identity 0 = γ(t, ·)# L 1 ({0}) = L 1 ({x0 : γ(t, x0 ) = 0})
92
6. THE CONNECTION BETWEEN PDE AND ODE
and use Fubini’s theorem to obtain Z L 1 ({t : γ(t, x0 ) = 0}) dx0 . 0= R
L 1 -a.e.
Hence, for x0 , the trajectory γ(·, x0 ) does not stay at 0 for a strictly positive set of times. This example shows that we cannot always expect stability: we have exhibited a family of smooth approximating sequences which converge toward a family of distinct flows of the equation. However, only one of these flows satisfies the compressibility assumption. The point here is the lack of control in L∞ on the divergence of the vector field (compare with Theorem 6.12 and with the first example in Section 5.3). 4. Ambrosio’s theory of regular Lagrangian flows In this section we present the derivation of the results of existence, uniqueness and stability for the regular Lagrangian flow deduced from the well-posedness in the class of bounded solutions for the continuity equation. This abstract passage is due to Ambrosio; we refer to [8] for the original approach in the BV case and to the lecture notes [9], [10] and [13] for the formalization of the argument in the general case. We notice that these results, in the Sobolev framework, have been already obtained by DiPerna and Lions in [83] with different techniques. We also mention the recent work by Figalli [92] in which a similar theory is developed in the context of stochastic differential equations. This approach is strongly based on the notion of superposition solution: starting from the “generalized flow” given by the measures ηx ∈ P(Rd ) we perform various constructions at the level of measure-valued solutions of the continuity equation; at that point the PDE well-posedness comes into play, allowing to deduce results about the measures ηx , roughly speaking obtaining that the generalized flow is in fact a regular Lagrangian flow, since ηx selects a single trajectory for L d -a.e. x ∈ Rd . Theorem 6.7 (Existence and uniqueness of the regular Lagrangian flow). Let b : [0, T ] × Rd → Rd be a bounded vector field. Assume that the continuity equation (6.1) has the uniqueness property in L∞ ([0, T ] × Rd ). Then the regular Lagrangian flow associated to b, if it exists, is unique. Assume in addition that the continuity equation (6.1) with initial data µ ¯ = L d has a positive solution in L∞ ([0, T ] × Rd ). Then we have existence of a regular Lagrangian flow relative to b. With a little abuse of terminology, in the statement of the theorem we write “µt ∈ meaning that we require that µt is an absolutely continuous (with respect to L d ) measure, whose density is essentially uniformly bounded. With uniqueness property in L∞ ([0, T ] × Rd ) we mean that, for every initial data µ ¯ ∈ L∞ (Rd ), if µ1t , µ2t ∈ ∞ d 1 2 L ([0, T ] × R ) are two solutions of (6.1) with µ0 = µ0 = µ ¯, then we must have 1 2 d+1 d µt (x) = µt (x) for L -a.e. (t, x) ∈ [0, T ] × R . Notice that, using a simple regularization argument, the existence of a positive solution assumed in the theorem is satified for instance if b has bounded divergence.
L∞ ”
Remark 6.8 (Uniqueness of Lb -Lagrangian flows). Theorem 6.7 could be generalized as follows, recalling the notion of Lb -Lagrangian flow discussed in Remark 6.5:
4. AMBROSIO’S THEORY OF REGULAR LAGRANGIAN FLOWS
93
if the continuity equation has the uniqueness property in Lb , then there is a unique Lb -Lagrangian flow. The proof is very similar to the one we are going to present: the reader is referred to Section 4 of [13] for the details and the modifications needed. In the remaining part of this section we illustrate the proof of Theorem 6.7. We first show that Theorem 6.7 is a consequence of the following proposition. Proposition 6.9. Let b : [0, T ] × Rd → Rd be a bounded vector field. Assume that the continuity equation (6.1) has the uniqueness property in L∞ ([0, T ]×Rd ). Consider a family {ηx }x∈Rd ⊂ P(ΓT ) such that ηx is concentrated on absolutely continuous integral solutions of the ordinary differential equation starting from x, for L d -a.e. x ∈ Rd . Assume that the superposition solution of the continuity equation µηt x induced by this family belongs to L∞ ([0, T ] × Rd ). Then ηx is a Dirac mass for L d -a.e. x ∈ Rd . Remark 6.10. We notice that the uniqueness result of Proposition 6.9 is in a certain sense stronger than the one of Theorem 6.7, since it holds in the wider class of the “multivalued solutions” given by the measures {ηx }x∈Rd . Proof of Theorem 6.7. Uniqueness of the regular Lagrangian flow. Assume that there exist two different regular Lagrangian flows X1 (t, x) and X2 (t, x) relative to the vector field b. We show that it is possibile to construct a family of measures {ηx }x∈Rd ⊂ P(ΓT ) such that • for L d -a.e. x ∈ Rd the measure ηx is concentrated on absolutely continuous integral solutions of the ordinary differential equation starting from x; • the superposition solution µηt x induced by the family {ηx }x∈Rd belongs to L∞ ([0, T ] × Rd ); • the measure ηx is not a Dirac mass for every x belonging to a set of positive Lebesgue measure. These three properties are clearly in contrast with the result in Proposition 6.9: hence, if we are able to construct such a family, we obtain a contradiction that gives the proof of the theorem. We first consider for L d -a.e. x ∈ Rd the measures ηx1 = δX1 (·,x) and ηx2 = δX2 (·,x) . Define 1 1 ηx + ηx2 . ηx = 2 We check that the superposition solution µηt x is a bounded function. Indeed for every ϕ ∈ Cc (Rd ) we can compute Z Z ηx hµt , ϕi = ϕ(γ(t)) dηx (γ) dx Rd ΓT Z Z 1 δ + δX2 (·,x) dx = ϕ(γ(t)) d 2 X1 (·,x) Rd ΓT Z 1 = ϕ(X1 (t, x)) + ϕ(X2 (t, x)) dx Rd 2 Z Z 1 1 d = ϕ(y) d X1 (t, ·)# L (y) + ϕ(y) d X2 (t, ·)# L d (y) . 2 Rd 2 Rd
94
6. THE CONNECTION BETWEEN PDE AND ODE
This means that 1 X1 (t, ·)# L d + X2 (t, ·)# L d , 2 and by condition (ii) in the definition of regular Lagrangian flow we obtain that µηt x belongs to L∞ ([0, T ] × Rd ). The fact that ηx is concentrated on solutions of the ODE is clear by its definition. Moreover, ηx is not a Dirac mass for every x ∈ Rd such that X1 (·, x) 6= X2 (·, x), and this happens in a set of points x of positive measure, since we are assuming that the two regular Lagrangian flows X1 and X2 are different. Existence of the regular Lagrangian flow. Let µt ∈ L∞ ([0, T ] × Rd ) be a positive solution of the continuity equation with initial data µ ¯ = L d . Noticing that the assumptions of Theorem 6.3 are satisfied, we apply the superposition principle deducing that µt = µηt x for some family {ηx }x∈Rd , with ηx concentrated on absolutely continuous integral solutions of the ODE starting from x, for L d -a.e. x ∈ Rd . Since the continuity equation has the uniqueness property in L∞ ([0, T ] × Rd ) we apply Proposition 6.9 to deduce that ηx is a Dirac mass for L d -a.e. x ∈ Rd . Denote by X(·, x) the element of ΓT on which ηx is concentrated, for L d -a.e. x ∈ Rd ; we check that the map X(t, x) defined in this way is a regular Lagrangian flow associated to b. Condition (i) in Definition 6.4 is clearly satisfied because ηx is concentrated on solutions of the ODE. To check condition (ii) it is enough to notice that X(t, ·)# L d = µt ∈ L∞ ([0, T ] × Rd ). µηt x =
We now pass to the proof of Proposition 6.9. We will use the following general criterion, whose proof is rather easy. Lemma 6.11. Let {ηx }x∈Rd ⊂ P(ΓT ) satisfy the following property: for every t ∈ [0, T ] and every pair of disjoint Borel sets E1 , E2 ⊂ Rd we have ηx {γ : γ(t) ∈ E1 } ηx {γ : γ(t) ∈ E2 } = 0 for L d -a.e. x ∈ Rd . Then ηx is a Dirac mass for L d -a.e. x ∈ Rd . Proof of Proposition 6.9. We argue by contradiction and we use the criterion in Lemma 6.11. We find t¯ ∈]0, T ], a Borel set C ⊂ Rd with L d (C) > 0 and a couple of disjoint Borel sets E1 , E2 ⊂ Rd such that ηx {γ : γ(t¯) ∈ E1 } ηx {γ : γ(t¯) ∈ E2 } = 0 for every x ∈ C. Possibly passing to a smaller set C still having strictly positive Lebesgue measure we can assume that (6.16) 0 < ηx {γ : γ(t¯) ∈ E1 } ≤ M ηx {γ : γ(t¯) ∈ E2 } for every x ∈ C for some constant M . We now want to localize to trajectories starting from the set C and arriving (at time t¯) in the sets E1 and E2 . We define ηx1 = 1C (x)ηx {γ : γ(t¯) ∈ E1 }
and
ηx2 = M 1C (x)ηx {γ : γ(t¯) ∈ E2 } .
Now denote by µ1t and µ2t (for t ∈ [0, t¯]) the superposition solutions of the continuity equation induced by the families of measures ηx1 and ηx2 respectively. Notice that the superposition solutions are well-defined even if ηx1 and ηx2 are not probability measures,
4. AMBROSIO’S THEORY OF REGULAR LAGRANGIAN FLOWS
95
but just positive measures with finite mass for L d -a.e. x ∈ Rd . For every ϕ ∈ Cc (Rd ) we can compute Z Z 1 hµ0 , ϕi = ϕ(γ(0)) d 1C (x)ηx {γ : γ(t¯) ∈ E1 } (γ) dx Rd Γ Z Z T = ϕ(γ(0)) dηx (γ) dx C {γ : γ(t¯)∈E1 } Z ϕ(x)ηx {γ : γ(t¯) ∈ E1 } dx , = C
from which we deduce µ10 = ηx {γ : γ(t¯) ∈ E1 } L d C . An analogous computation gives µ20 = M ηx {γ : γ(t¯) ∈ E2 } L d C . Recalling (6.16) we obtain that µ10 ≤ µ20 . Now let f : Rd → [0, 1] be the density of µ10 with respect to µ20 (i.e. f satisfies µ10 = f µ20 ) and set η˜x2 = M f (x)1C (x)ηx {γ : γ(t¯) ∈ E2 } . Consider the superposition solution µ ˜2t (defined for t ∈ [0, t¯]) induced by the family of ˜20 and that measures η˜x2 . We can readily check that µ10 = µ Z Z hµ1t¯ , ϕi = ϕ(γ(t¯)) dηx (γ) dx C
and h˜ µ2t¯ , ϕi
{γ : γ(t¯)∈E1 }
Z =
Z M f (x)
C
ϕ(γ(t¯)) dηx (γ) dx .
{γ : γ(t¯)∈E2 } E1 and µ ˜2t¯ is
We deduce that µ1t¯ is concentrated on concentrated on E2 . Hence µ1t and 2 ∞ d µ ˜t are solutions in L ([0, T ] × R ) of the continuity equation with the same initial data at time t = 0, but they are different at time t = t¯. We are violating the uniqueness assumption and from this contradiction we obtain the thesis. In a similar fashion it is possible to show a result of abstract stability: if the continuity equation is well-posed for each approximating vector field bk then the associated regular Lagrangian flows converge strongly to the regular Lagrangian flow associated to the limit vector field b. Theorem 6.12 (Stability of the regular Lagrangian flow). Consider a sequence {bk } of vector fields such that (6.17)
kbk kL∞ ([0,T ]×Rd ) + kdiv bk kL∞ ([0,T ]×Rd ) ≤ C
and assume that for each bk the continuity equation has the uniqueness property in L∞ ([0, T ] × Rd ). Assume that the sequence {bk } converges in L1loc ([0, T ] × Rd ) to a vector field b ∈ L∞ ([0, T ]×Rd ) with div b ∈ L∞ ([0, T ]×Rd ). Assume that the continuity equation with vector field b has the uniqueness property in L∞ ([0, T ] × Rd ). Then the regular Lagrangian flows Xk associated to bk converge strongly in L∞ ([0, T ]; L1loc (Rd )) to the regular Lagrangian flow X associated to b.
96
6. THE CONNECTION BETWEEN PDE AND ODE
We only give a sketch of the proof, which is closely related to the one of the uniqueness result of Theorem 6.7. We refer the reader to Section 3 of [10] for more details. As in the proof of the superposition principle we consider for every k the measure η k ∈ M+ (Rd × ΓT ) associated to the regular Lagrangian flow X k . Using the bounds (6.17) it is possible to show that η k is locally tight. Consider a limit point η ∈ M+ (Rd × ΓT ). Arguing as in Step 6 of the proof of Theorem 6.3 and using the convergence of bk to b we obtain that η is concentrated on trajectories of the ODE with vector field b. Moreover from the narrow convergence and from (6.17) we get that η induces bounded superposition solutions to the continuity equation. We can then apply Proposition 6.9 to get that η is in fact concentrated on a graph of a regular Lagrangian flow associated to b. But under our assumptions we already know from Theorem 6.7 that such a regular Lagrangian flow is unique, and this means that we have (x, X k (·, x))# L d → (x, X(·, x))# L d locally narrowly. From this it is possible to deduce the strong convergence of X k to X in L∞ ([0, T ]; L1loc (Rd )). Remark 6.13. The strength of this abstract approach relies in the fact that, having at our disposal the PDE well-posedness results for Sobolev or BV vector fields (see Sections 2.5 and 2.6), for the various cases addressed in Section 2.7 and for the twodimensional case (see Chapter 4) we can immediately conclude, using Theorems 6.7 and 6.12, existence, uniqueness and stability of the regular Lagrangian flow in all these situations (see also Remarks 2.8 and 2.9). 5. DiPerna–Lions’ notion of flow The approach to the ordinary differential equation due to DiPerna and Lions [83] is quite different. It is based on an observation that we already made in the smooth setting: the flow X(t, s, x) starting at time s from the point x satisfies ∂X (t, s, x) + b(s, x) · ∇x X(t, s, x) = 0 ∂s (see equation (1.15)), that is a transport equation with vector field b. In this way results relative to the PDE can be transferred to the ODE. DiPerna–Lions’ notion of flow is based on the following three axioms: (a) The ordinary differential equation ∂ X(t, x) = b(t, X(t, x)) ∂t X(0, x) = x is satisfied in the sense of distributions in [0, T ] × Rd ; (b) The push-forward of the Lebesgue measure satisfies 1 d L ≤ X(t, ·)# L d ≤ CL d ; C
6. THE DENSITY OF A REGULAR LAGRANGIAN FLOW
97
(c) The semi-group property holds: for L d -a.e. x ∈ Rd we have X(t, X(s, x)) = X(s + t, x)
for every s, t ≥ 0 with s + t ≤ T .
Up to a redefinition of the flow X on a negligible set, condition (a) is equivalent to condition (i) in the definition of regular Lagrangian flow. However, Definition 6.4(ii) only asks for an upper bound on the push-forward of the Lebesgue measure, while (b) also requires a bound from below. Moreover, the semi-group property (c) is completely removed in the approach of [8]: it is in fact a consequence of the other two assumptions, using the following argument (see Remark 6.7 of [8]). Fix s ∈ [s0 , T ] and define X(t, s0 , x) if t ∈ [s0 , s] ˜ X(t, x) = 0 X(t, s, X(s, s , x)) if t ∈ [s, T ]. ˜ x) is a solution of the ODE in [s0 , T ] for L d -a.e. x ∈ It is immediate to check that X(·, ˜ Rd . Moreover we easily obtain the following compressibility condition for X: ˜ ·)# L d ≤ L2 L d , X(t, where L is a compressibility constant for the regular Lagrangian flow X. By the ˜ x) = X(t, s0 , x), and this means that the semiuniqueness result we obtain that X(t, group property (c) holds. In this framework we also mention the recent result by Hauray, Le Bris and Lions [100] in which uniqueness of this notion of flow (for Sobolev vector fields) is shown directly at the ODE level, with a very simple argument. 6. The density of a regular Lagrangian flow We now present another approach to the theory of regular Lagrangian flows, based on the concept of density transported by the flow and particularly useful to deal with nearly incompressible vector fields (recall Definition 2.14); see Section 2.8 for the PDE theory in this framework. The presentation here is very close to the one by De Lellis ([76] and [78]) and we refer to these papers for a complete exposition. This approach works naturally under a compressibility assumption which is a bit weaker than the one assumed in the standard definition of regular Lagrangian flow (see Definition 6.4(ii)): we will only need that (6.18)
X(t, ·)# L d L d
for L 1 -a.e. t ∈ [0, T ].
If X is a regular Lagrangian flow associated to a bounded vector field b we define µX = id , X # L d+1 [0, T ] × Rd , i.e. µX is the push-forward of the Lebesgue measure on [0, T ] × Rd via the map (t, x) 7→ (t, X(t, x)). The compressibility condition (6.18) gives the existence of a function ρ ∈ L1loc ([0, T ] × Rd ) such that (6.19)
µX = ρL d+1 ([0, T ] × Rd ) .
98
6. THE CONNECTION BETWEEN PDE AND ODE
As already remarked in Section 2.8, in the case of a smooth vector field we can explicitly compute ρ(t, x) = det ∇x X t, X(t, ·)−1 (x) . Definition 6.14. The function ρ defined via (6.19) is called the density of the regular Lagrangian flow X. The following proposition establishes a link between this notion and the Eulerian side of the problem, namely with the continuity equation relative to the vector field b. Proposition 6.15. Let X be a regular Lagrangian flow associated to a bounded nearly incompressible vector field b and let ζ¯ ∈ L∞ (Rd ). Define ¯ d+1 [0, T ] × Rd . µ = id , X # ζL Then there exists ζ ∈ L1loc ([0, T ] × Rd ) such that µ = ζL d+1 ([0, T ] × Rd ); moreover ζ is a solution of ( ∂t ζ + div (bζ) = 0 (6.20) ζ(0, ·) = ζ¯ . The density of the regular Lagrangian flow corresponds to the case ζ¯ = 1 in the previous proposition: this means that the density ρ satisfies ( ∂t ρ + div (bρ) = 0 ρ(0, ·) = 1 . In particular we deduce that, when b is a bounded nearly incompressible vector field satisfying the renormalization property (recall Definition 2.15), the density of a regular Lagrangian flow X relative to b coincides with the density generated by b in the sense of Definition 2.18. This is precisely the link between the Lagrangian and the Eulerian problems: the proof of the following theorem, relative to the well-posedness for the regular Lagrangian flow, is strongly based on the well-posedness for the continuity equation in Theorem 2.17. Theorem 6.16. Assume that b is a bounded nearly incompressible vector field such that the extension ˜b as in (2.36) has the renormalization property in the sense of Definition 2.15. Then there exists a unique regular Lagrangian flow X associated to b. Moreover, let {bk } be a sequence of bounded nearly incompressible vector fields such that the extensions ˜bk as in (2.36) have the renormalization property in the sense of Definition 2.15 and, for every k, let ρk be as in (2.35). Assume that there exists a constant C such that kbk k∞ + kρk k∞ + kρ−1 k k∞ ≤ C and that bk → b in L1loc ([0, T ] × Rd ). Then the regular Lagrangian flows Xk associated to bk converge strongly in L1loc ([0, T ] × Rd ) to X. Proof. The existence part is essentially based on an approximation procedure, which however requires some care, and for this we refer to Theorem 3.22 of [76].
6. THE DENSITY OF A REGULAR LAGRANGIAN FLOW
99
We now show uniqueness. Let X and Y be two regular Lagrangian flows relative to the vector field b. Fix ζ¯ ∈ Cc (Rd ) and, recalling Theorem 2.17, consider the unique solution ζ of (6.20). According to Proposition 6.15 we have ¯ d+1 [0, T ] × Rd ζL d+1 [0, T ] × Rd = id , X # ζL ¯ d+1 [0, T ] × Rd . = id , Y # ζL This identity means Z Z TZ ¯ dx dt = ϕ(x, X(t, x))ζ(x) Rd
0
T
Z
¯ dx dt ϕ(x, Y (t, x))ζ(x)
Rd
0
for every ϕ ∈ Cc ([0, T ] × Rd ). Since ζ¯ has compact support this identity is also valid when ϕ(t, y) = χ(t)yj , for χ ∈ C([0, T ]) and j = 1, . . . , d. Then Z TZ Z TZ ¯ dx dt ¯ dx dt = Yj (t, x)χ(t)ζ(x) Xj (t, x)χ(t)ζ(x) Rd
0
Rd
0
for any pair of functions χ ∈ C([0, T ]) and ζ¯ ∈ Cc (Rd ). This implies that X(t, x) = Y (t, x) for L d+1 -a.e. (t, x) ∈ [0, T ] × Rd . We finally show the stability property. Let bk → b as in the statement of the theorem. Fix a function ζ¯ ∈ Cc (Rd ). Applying Theorem 2.17 we can define ζ and ζk as the unique solutions to the problems ( ( ∂t ζ + div (bζ) = 0 ∂t ζk + div (bk ζk ) = 0 and ζ(0, ·) = ζ¯ ζk (0, ·) = ζ¯ . By the weak stability of the continuity equation we deduce that, up to subsequences, ∗ ζk * ζ in L∞ ([0, T ] × Rd ) − w∗ . Recalling the characterization in Proposition 6.15 we see that this means Z TZ Z TZ ¯ dtdx → ¯ dtdx (6.21) ϕ t, Xk (t, x) ζ(x) ϕ t, X(t, x) ζ(x) Rd
0
Rd
0
for every ϕ ∈ Cc ([0, T ] × Rd ). However, noticing that the regular Lagrangian flows Xk ¯ are locally equi-bounded, we are allowed to use as test functions ζ(x) = 1BR (0) (x) and 2 ϕ(t, x) = |x| , and this gives Z TZ Z TZ 2 (6.22) |Xk (t, x)| dtdx → |X(t, x)|2 dtdx . 0
BR (0)
0
BR (0)
¯ For the same reason we can substitute in (6.21) ζ(x) = β(x)1BR (0) (x) and ϕ(t, x) = d d γ(t)x · v, where β ∈ Cc (R ), γ ∈ C([0, T ]) and v ∈ R . Since we can approximate X strongly in L1loc ([0, T ] × Rd ) with functions of the form N X
vi γi (t)βi (x)
i=1
we obtain Z
T
Z
Z
T
Z
Xk (t, x) · X(t, x) dtdx →
(6.23) 0
BR (0)
0
BR (0)
|X(t, x)|2 dtdx .
100
6. THE CONNECTION BETWEEN PDE AND ODE
Putting together (6.22) and (6.23) we get the desired conclusion.
7. Bressan’s compactness conjecture In [48] Bressan proposed the following conjecture. Conjecture 6.17 (Bressan’s compactness conjecture). Let bk : [0, T ] × Rd → Rd , k ∈ N, be a sequence of smooth vector fields and denote by Xk the (classical) flows associated to them, i.e. the solutions of ∂Xk (t, x) = bk (t, Xk (t, x)) ∂t (6.24) Xk (0, x) = x . Assume that kbk k∞ +k∇bk kL1 is uniformly bounded and that the flows Xk are uniformly nearly incompressible, i.e. that 1 (6.25) ≤ det(∇x Xk (t, x)) ≤ C for some constant C > 0. C Then the sequence {Xk } is strongly precompact in L1loc ([0, T ] × Rd ). This conjecture was advanced in connection with the Keyfitz and Kranzer system, in particular to provide the existence of suitable weak solutions (recall Section 5.2). In fact it is possible to prove well-posedness results for this system bypassing Conjecture 6.17 (see [15] and [11]), which nevertheless remains an interesting question. In this section we indicate some advances in its proof, mainly contained in [16]. In [11] it has been proved that Conjecture 6.17 would follow from the following one. Conjecture 6.18. Any bounded nearly incompressible BV vector field has the renormalization property in the sense of Definition 2.15. Conjecture 6.18 implies Conjecture 6.17. Let ρk be the density generated by Xk . From (6.25) it follows the existence of a constant C˜ > 0 such that 1 ≤ ρk ≤ C˜ . C˜ From the BV compactness theorem and the weak∗ compactness of L∞ it is sufficient to prove Conjecture 6.17 under the assumptions that bk → b strongly in L1loc ([0, T ] × Rd ) ∗ for some BV vector field b and that ρk * ρ in L∞ ([0, T ] × Rd ) − w∗ for some bounded function ρ. We notice that this implies that b is bounded, ρ ≥ 1/C˜ and ∂t ρ + div (bρ) = 0 . The last identity can be deduced passing to the limit in ∂t ρk + div (bk ρk ) = 0. Then b is a bounded nearly incompressible BV vector field and, if Conjecture 6.18 has a positive answer, b has the renormalization property. If this is the case, we apply the stability result in Theorem 6.16 to conclude that Xk converges strongly in L1loc ([0, T ] × Rd ) to the unique regular Lagrangian flow associated to b.
7. BRESSAN’S COMPACTNESS CONJECTURE
101
In order to present the main results of [16] we need to introduce some notation, taken from [14]. Let Ω ⊂ Rm be an open set and consider B ∈ BVloc (Ω; Rm ). Consider ˜ consisting of the points x ∈ Ω such that the limit the set E DB(Br (x)) M (x) = lim r→0 |DB|(Br (x)) ˜ exists and is finite and the Lebesgue limit B(x) exists. The tangential set of B is defined as follows: n o ˜ : M (x) · B(x) ˜ (6.26) E= x∈E =0 . We can think of this set as the set on which the vector field B is “orthogonal to its derivative”. Building on some results of [14], in [16] the following result is shown. Theorem 6.19. Let b : [0, T ] × Rd → Rd be a bounded nearly incompressible BV vector field. Consider the vector field B = (1, b) : [0, T ] × Rd → R × Rd and let E be the tangential set of B. If (6.27)
c |Dt,x · B|(E) = |Dxc · b|(E) = 0
then b has the renormalization property. c · B and D c · b for the In the statement of the theorem we used the notation Dt,x x Cantor part of the space-time and of the space divergence of B and b respectively. The presence of condition (6.27) in the theorem leads to the following question. m Question 6.20. Let B ∈ BVloc ∩ L∞ loc (Ω; R ) and denote by E its tangential set. c Under which conditions does the equality |D · B|(E) = 0 hold?
Indeed some conditions are needed, as shown by a counterexample presented in Section 8 of [16]. Notice that a positive answer to the following question would be enough for the proof of Conjecture 6.18; however, the answer to this second question is presently not known. m Question 6.21. Let B ∈ BVloc ∩ L∞ loc (Ω; R ) and denote by E its tangential set. Assume that there exists a function ρ ∈ L∞ (Ω) such that ρ ≥ C > 0 and div (Bρ) = 0. Then does the equality |Dc · B|(E) = 0 hold?
Another approach to Bressan’s compactness conjecture, based on quantitative a priori estimates for regular Lagrangian flows (and hence bypassing the connection with the theory of renormalized solutions), will be presented in the next chapter.
CHAPTER 7
A priori estimates for regular Lagrangian flows In this chapter we present a joint work with De Lellis [66] (see also Section 8 of [13] and [67] for another exposition of these results). We are able to show some quantitative estimates for W 1,p vector fields, with p > 1, which allow to recover the results of existence, uniqueness and stability of regular Lagrangian flows presented in the previous chapter. Moreover these estimates have some new and interesting corollaries, regarding compactness, quantitative regularity and quantitative stability of regular Lagrangian flows. They also allow to prove a propagation of a mild regularity for solutions to the transport equation. One of the merits of this approach is the fact that the whole derivation is purely Lagrangian, in the sense that everything is deduced from the estimates and from the definition of regular Lagrangian flow only, with no mention to the Eulerian side of the problem (compare with the derivation we presented in Section 6.4). The only drawback lies in the assumption p > 1, which can be relaxed a bit (in fact our arguments work under the assumption Db ∈ L log L, see Sections 7.2 and 7.4), but we are presently not able to cover the case p = 1, and this does not allow to reach the very important BV setting. 1. A purely Lagrangian approach We recall that, as we discussed in Section 6.3, for a vector field b which is merely locally summable we can give the following definition of regular Lagrangian flow. This notion turns out to be the right one in the study of the ordinary differential equation with weakly differentiable vector field. Notice that the boundedness assumption on the vector field assumed in the previous chapters is not essential: in this chapter we will also address the theory of regular Lagrangian flows relative to vector fields satisfying more general growth conditions. Definition 7.1 (Regular Lagrangian flow). Let b ∈ L1loc ([0, T ] × Rd ; Rd ). We say that a map X : [0, T ] × Rd → Rd is a regular Lagrangian flow for the vector field b if (i) for L d -a.e. x ∈ Rd the map t 7→ X(t, x) is an absolutely continuous integral solution of γ(t) ˙ = b(t, γ(t)) for t ∈ [0, T ], with γ(0) = x; (ii) there exists a constant L independent of t such that (7.1)
X(t, ·)# L d ≤ LL d .
The constant L in (ii) will be called the compressibility constant of X. We illustrated in the previous chapter the results by DiPerna and Lions [83] and by Ambrosio [8] regarding existence, uniqueness and stability of regular Lagrangian flows, in the Sobolev and BV context respectively. However the argument that we 103
104
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
have shown is quite indirect: as we have seen it exploits the connection between the ordinary differential equation and the Cauchy problem for the continuity equation. We show in this chapter how many of the results of the DiPerna–Lions theory relative to regular Lagrangian flows can be recovered from simple a priori estimates, directly in the Lagrangian formulation. Though our approach works under various relaxed hypotheses, namely controlled growth at infinity of the field b and Lploc and L log L assumptions on Dx b, for simplicity in this introductory discussion we consider a vector field b in W 1,p ∩ L∞ , with p > 1. Assuming the existence of a regular Lagrangian flow X, we give estimates of integral quantities depending on X(t, x) − X(t, y). These estimates depend only on kbkW 1,p + kbk∞ and the compressibility constant L of Definition 7.1(ii). Moreover, a similar estimate can be derived for the difference X(t, x) − X 0 (t, x) of regular Lagrangian flows of different vector fields b and b0 , depending only on the compressibility constant of b and on kbkW 1,p + kbk∞ + kb0 k∞ + kb − b0 kL1 . As direct corollaries of our estimates we then derive: (a) Existence, uniqueness, stability, and compactness of regular Lagrangian flows; (b) Some mild regularity properties, like the approximate differentiability proved in [19], that we recover in a new quantitative fashion. The regularity property in (b) has an effect on solutions to the transport equation: we can prove that, for b ∈ W 1,p ∩ L∞ with bounded divergence, solutions to the transport equation propagate the same mild regularity of the corresponding regular Lagrangian flow (we refer to Section 7.5 for the precise statements). Our approach has been inspired by a recent result of Ambrosio, Lecumberry and Maniglia [19], proving the almost everywhere approximate differentiability of regular Lagrangian flows. Indeed, some of the quantities we estimate in this paper are taken directly from [19], whereas others are just suitable modifications. However, the way we derive our estimates is different: our analysis relies all on the Lagrangian formulation, whereas that of [19] relies on the Eulerian one. We also mention a previous work by Le Bris and Lions [103], in which, among other things, a kind of “differentiability in measure” of the flow is shown. The relation between this notion and the classical approximate differentiability has been studied by Ambrosio and Mal´ y [21]. See also Sections 6 and 7 of [13] for an account of these results. Unfortunately with our approach we do not recover all the results of the theory of renormalized solutions. The main problem is that our estimates do not conver the case Db ∈ L1 . Actually, the extension to the case Db ∈ L1 of our (or of similar) estimates would answer positively to the compactness conjecture made by Bressan in [48] (see Conjecture 6.17). At the present stage, the theory of renormalized solutions cannot be extended to cover this interesting case (see Section 6.7 for an account of some partial results in this direction). In another paper, [47], Bressan raised a second conjecture on mixing properties of flows of BV vector fields (see Conjecture 7.33 below), which can be considered as a quantitative version of Conjecture 6.17. In Section 6 we show how our estimates settle the W 1,p (p > 1) analog of Bressan’s mixing conjecture. In order to keep the presentation simple, in Section 7.2 we give the estimates and the various corollaries in the case b ∈ W 1,p ∩ L∞ and in Section 7.3 we present the more general estimates and their consequences. We thank Herbert Koch for suggesting us
2. A PRIORI ESTIMATES FOR BOUNDED VECTOR FIELDS AND COROLLARIES
105
that the Lipschitz estimates hold under the assumption Db ∈ L log L (see Remark 7.5 and the discussion at the beginning of Section 7.4). In Section 7.4 we show how to prove directly, via suitable a priori estimates, the compactness conclusion of Conjecture 6.17 when Dbk is bounded in L log L. It has been pointed out independently by Fran¸cois Bouchut and by Pierre-Emmanuel Jabin that a more careful analysis allows to extend this approach when the sequence {Dbk } is equi-integrable. In Section 7.5 we discuss the regularity results for transport equations mentioned above. Finally, in Section 7.6 we prove the W 1,p analog of Bressan’s mixing conjecture. In all the chapter constants will be denoted by c and ca1 ,...,aq , where we understand that in the first case the constant is universal and in the latter that it depends only on the quantities a1 , . . . , aq . Therefore, during several computations, we will use the same symbol for constants which change from line to line. 2. A priori estimates for bounded vector fields and corollaries In this section we show our estimates in the particular case of bounded vector fields. This estimate and its consequences are just particular cases of the more general theorems presented in the next sections. However, we decided to give independent proofs in this simplified setting in order to illustrate better the basic ideas of our analysis. 2.1. Estimate of an integral quantity and Lipschitz estimates. Theorem 7.2. Let b be a bounded vector field belonging to L1 ([0, T ]; W 1,p (Rd ; Rd )) for some p > 1 and let X be a regular Lagrangian flow associated to b. Let L be the compressibility constant of X, as in Definition 7.1(ii). For every p > 1 define the following integral quantity: "Z !p #1/p Z |X(t, x) − X(t, y)| + 1 dy dx . log sup sup − Ap (R, X) = r BR (0) 0≤t≤T 0 1, or equi-bounded in L∞ ([0, T ] × Rd ; Rd ) and in L1 ([0, T ]; W 1,1 (Rd ; Rd )) and such that {Mλ Dbh } is equi-bounded in L1 ([0, T ]; L1 (Rd )) for every λ > 0. For each h, let Xh be a regular Lagrangian flow associated to bh and let Lh be the compressibility constant of Xh , as in Definition 7.1(ii). Suppose that the sequence {Lh } is equi-bounded. Then the sequence {Xh } is strongly precompact in L1loc ([0, T ] × Rd ). Proof. Fix δ > 0 and R > 0. Since {bh } is equi-bounded in L∞ ([0, T ] × Rd ), we deduce that {Xh } is equi-bounded in L∞ ([0, T ] × BR (0)): let C1 (R) be an upper bound for these norms. Applying Proposition 7.4, for every h we find a Borel set Kh,δ such that L d (BR (0) \ Kh,δ ) ≤ δ and cd Ap (R, Xh ) Lip Xh (t, ·)|Kh,δ ≤ exp for every t ∈ [0, T ]. δ 1/p Recall first Theorem 7.2 implies that Ap (R, Xh ) is equi-bounded with respect to h, because of the assumptions of the corollary. Moreover, using Definition 7.1(i) and thanks again to the equi-boundedness of {bh } in L∞ ([0, T ] × Rd ), we deduce that there exists a constant C2δ (R) such that Lip Xh |[0,T ]×Kh,δ ≤ C2δ (R) . If we now set Bh,δ = [0, T ]×Kh,δ and Mδ = max C1 (R), C2δ (R) , we are in the position to apply Lemma 7.7 with Ω = [0, T ] × BR (0). Then the sequence {Xh } is precompact in measure in [0, T ] × BR (0), and by equi-boundedness in L∞ we deduce that it is also precompact in L1 ([0, T ] × BR (0)). Using a standard diagonal argument it is possible to conclude that {Xh } is locally precompact in L1 ([0, T ] × Rd ). Corollary 7.9 (Existence of the flow). Let b be a bounded vector field belonging to L1 ([0, T ]; W 1,p (Rd ; Rd )) for some p > 1, or belonging to L1 ([0, T ]; W 1,1 (Rd ; Rd )) and satisfying Mλ Db ∈ L1 ([0, T ]; L1 (Rd )) for every λ > 0. Assume that
2. A PRIORI ESTIMATES FOR BOUNDED VECTOR FIELDS AND COROLLARIES
111
[div b]− ∈ L1 ([0, T ]; L∞ (Rd )). Then there exists a regular Lagrangian flow associated to b. Proof. This is a simple consequence of the previous corollary. Choose a positive convolution kernel in Rd and regularize b by convolution. It is simple to check that the sequence of smooth vector fields {bh } we have constructed satisfies the equi-bounds of the previous corollary. Moreover, since every bh is smooth, for every h there is a unique regular Lagrangian flow associated to bh , with compressibility constant Lh given by Z (7.8)
T
Lh = exp
k[div bh (t, ·)]− kL∞ (Rd ) dt .
0
Thanks to the positivity of the chosen convolution kernel, the sequence {Lh } is equibounded, then we can apply Corollary 7.8. It is then easy to check that every limit point of {Xh } in L1loc ([0, T ] × Rd ) is a regular Lagrangian flow associated to b. Remark 7.10. An analogous existence result could be obtained removing the hypothesis on the divergence of b, and assuming that there is some approximation procedure such that we can regularize b with equi-bounds on the compressibility constants of the approximating flows. This remark also applies to Corollaries 7.21 and 7.28. 2.3. Stability estimates and uniqueness. In this subsection we show an estimate similar in spirit to that of Theorem 7.2, but comparing flows for different vector fields. A direct corollary of this estimate is the stability (and hence the uniqueness) of regular Lagrangian flows. Theorem 7.11 (Stability of the flow). Let b and ˜b be bounded vector fields belonging ˜ be regular Lagrangian flows to L1 ([0, T ]; W 1,p (Rd ; Rd )) for some p > 1. Let X and X ˜ ˜ associated to b and b respectively and denote by L and L the compressibility constants of the flows. Then, for every time τ ∈ [0, T ], we have −1 ˜ kX(τ, ·) − X(τ, ·)kL1 (Br (0)) ≤ C log kb − ˜bkL1 ([0,τ ]×BR (0)) , ˜ where R = r + T kbk∞ and the constant C only depends on τ , r, kbk∞ , k˜bk∞ , L, L, and kDx bkL1 (Lp ) . Proof. Set δ = kb − ˜bkL1 ([0,τ ]×BR (0)) and consider the function Z g(t) =
log Br (0)
˜ x)| |X(t, x) − X(t, +1 δ
! dx .
112
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
Clearly g(0) = 0 and after some standard computations we get Z dX(t, x) dX(t, −1 ˜ x) ˜ x)| + δ g 0 (t) ≤ − dx |X(t, x) − X(t, dt dt Br (0) Z ˜ x))| |b(t, X(t, x)) − ˜b(t, X(t, = dx ˜ x)| + δ |X(t, x) − X(t, Br (0) Z 1 ˜ x)) − ˜b(t, X(t, ˜ x))| dx |b(t, X(t, ≤ δ Br (0) Z ˜ x))| |b(t, X(t, x)) − b(t, X(t, (7.9) + dx . ˜ x)| + δ |X(t, x) − X(t, Br (0) ˜ = 2r + T (kbk∞ + k˜bk∞ ) and we apply Lemma A.9 to estimate the last integral We set R as follows: Z Z ˜ x))| |b(t, X(t, x)) − b(t, X(t, ˜ x)) dx . dx ≤ cd MR˜ Db(t, X(t, x))+MR˜ Db(t, X(t, ˜ x)| + δ Br (0) |X(t, x) − X(t, Br (0) Inserting this estimate in (7.9), setting r˜ = r + T max{kbk∞ , k˜bk∞ }, changing variables in the integrals and using Lemma A.8 we get Z ˜Z L 0 ˜ ˜ |b(t, y) − b(t, y)| dy + L + L MR˜ Db(t, y) dy g (t) ≤ δ Br+T k˜bk (0) Br˜(0) ∞ ˜Z L ˜ + L kM ˜ Db(t, ·)kLp |b(t, y) − ˜b(t, y)| dy + cd r˜n−n/p L ≤ R δ Br+T k˜bk (0) ∞ ˜Z L ˜ + L kDb(t, ·)kLp . ≤ |b(t, y) − ˜b(t, y)| dy + cd,p r˜n−n/p L δ Br+T k˜bk (0) ∞
For any τ ∈ [0, T ], integrating the last inequality between 0 and τ we get ! Z ˜ x)| |X(τ, x) − X(τ, (7.10) g(τ ) = log + 1 dx ≤ C1 , δ Br (0) ˜ and kDx bkL1 (Lp ) . where the constant C1 depends on τ , r, kbk∞ , k˜bk∞ , L, L, Next we fix a second parameter η > 0 to be chosen later. Using the Chebyshev inequality we find a measurable set K ⊂ Br (0) such that L d (Br (0) \ K) ≤ η and ! ˜ x)| |X(τ, x) − X(τ, C1 +1 ≤ for x ∈ K. log δ η Therefore we can estimate Z ˜ x)| dx |X(τ, x) − X(τ, Br (0)
Z ˜ ˜ x)| dx ≤ η kX(τ, ·)kL∞ (Br (0)) + kX(τ, ·)kL∞ (Br (0)) + |X(τ, x) − X(τ,
K
(7.11) ≤ ηC2 + cd rd δ (exp(C1 /η)) ≤ C3 (η + δ exp(C1 /η)) ,
2. A PRIORI ESTIMATES FOR BOUNDED VECTOR FIELDS AND COROLLARIES
113
˜ and kDx bkL1 (Lp ) . with C1 , C2 and C3 which depend only on T , r, kbk∞ , k˜bk∞ , L, L, Without loss of generality we can assume δ < 1. Setting η = 2C1 | log δ|−1 = 2C1 (− log δ)−1 , we have exp(C1 /η) = δ −1/2 . Thus we conclude Z ˜ x)| dx ≤ C3 2C1 | log δ|−1 + δ 1/2 ≤ C| log δ|−1 , |X(τ, x) − X(τ, (7.12) Br (0)
˜ and kDx bkL1 (Lp ) . This completes the where C depends only on τ , r, kbk∞ , k˜bk∞ , L, L, proof. Corollary 7.12 (Uniqueness of the flow). Let b be a bounded vector field belonging to L1 ([0, T ]; W 1,p (Rd ; Rd )) for some p > 1. Then the regular Lagrangian flow associated to b, if it exists, is unique. Proof. It follows immediately from the stability proved in Theorem 7.11.
Remark 7.13 (Stability with weak convergence in time). Theorem 7.11 allows to show the stability when the convergence of the vector fields is just weak with respect to the time. This setting is in fact very natural in view of the applications to the theory of fluid mechanics (see Theorem II.7 in [83] and [106], in particular Theorem 2.5). In particular, under suitable bounds on the sequence {bh }, the following form of weak convergence with respect to the time is sufficient to get the thesis: Z T Z T bh (t, x)η(t) dt −→ b(t, x)η(t) dt in L1loc (Rd ) for every η ∈ Cc∞ (0, T ). 0
0
Indeed, fix a parameter > 0 and regularize with respect to the spatial variable only using a standard convolution kernel ρ . We can rewrite the difference Xh (t, x) − X(t, x) as Xh (t, x)−X(t, x) = Xh (t, x)−Xh (t, x) + Xh (t, x)−X (t, x) + X (t, x)−X(t, x) , where X and Xh are the flows relative to the regularized vector fields b and bh respectively. Now, it is simple to check that • The last term goes to zero with , by the classical stability theorem (the quantitative version is not needed at this point); • The first term goes to zero with , uniformly with respect to h: this is due to the fact that the difference bh − bh goes to zero in L1loc ([0, T ] × Rd ) uniformly with respect to h, if we assume a uniform control in W 1,p on the vector fields {bh }, hence we can apply Theorem 7.11, and we get the desired convergence; • The second term goes to zero for h → ∞ when is kept fixed, because we are dealing with flows relative to vector fields which are smooth with respect to the space variable, uniformly in time, and weak convergence with respect to the time is enough to get the stability. In order to conclude, we fix an arbitrary δ > 0 and we first find > 0 such that the norm of the third term is smaller than δ and such that the norm of the first term is smaller than δ for every h. For this fixed , we find h such that the norm of the
114
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
second term is smaller than δ. With this choice of h we have estimated the norm of Xh (t, x) − X(t, x) with 3δ, hence we get the desired convergence. Remark 7.14 (Another way to show compactness). If we apply Theorem 7.11 to ˜ x) = X(t, x + h) − h relative to the vector fields b(t, x) and the flows X(t, x) and X(t, ˜b(t, x) = b(t, x + h), where h ∈ Rd is fixed, we get for every τ ∈ [0, T ] −1 kX(τ, ·) − X(τ, · + h) − hkL1 (Br (0)) ≤ C log kb(t, x) − b(t, x + h)kL1 ([0,τ ]×BR (0)) C ≤ . | log(h)| Hence we have a uniform control on the translations in the space, and we can deduce a compactness result applying the Riesz-Fr´echet-Kolmogorov compactness criterion (Lemma 7.26). 3. Estimates for more general vector fields and corollaries In this section we extend the previous results to more general vector fields, in particular we drop the boundedness condition on b. More precisely, we will consider vector fields b : [0, T ] × Rd → Rd satisfying the following regularity assumptions: 1,p (Rd ; Rd )) for some p > 1; (R1) b ∈ L1 ([0, T ]; Wloc (R2) We can write b(t, x) = ˜b1 (t, x) + ˜b2 (t, x) 1 + |x|
with ˜b1 (t, x) ∈ L1 ([0, T ]; L1 (Rd ; Rd )) and ˜b2 (t, x) ∈ L1 ([0, T ]; L∞ (Rd ; Rd )). Since we are now considering vector fields which are no more bounded, we have to take care of the fact that the flow will be no more locally bounded in Rd . However, we can give an estimate of the measure of the set of the initial data such that the corresponding trajectories exit from a fixed ball at some time. Definition 7.15 (Sublevels). Fix λ > 0 and let X : [0, T ] × Rd → Rd be a locally summable map. We set (7.13) Gλ = x ∈ Rd : |X(t, x)| ≤ λ ∀t ∈ [0, T ] . Proposition 7.16 (Uniform estimate of the superlevels). Let b be a vector field satisfying assumption (R2) and let X be a regular Lagrangian flow associated to b, with compressibility constant L. Then we have L d (BR (0) \ Gλ ) ≤ g(R, λ) , where the function g only depends on k˜b1 kL1 (L1 ) , k˜b2 kL1 (L∞ ) and L; moreover we have that g(R, λ) ↓ 0 for R fixed and λ ↑ +∞. Proof. Let φt be the density of X(t, ·)# 1BR (0) L d with respect to L d and notice that, by the definition of push-forward and by Definition 7.1(ii), we have kφt k1 = ωd Rd
3. ESTIMATES FOR MORE GENERAL VECTOR FIELDS AND COROLLARIES
115
and kφt k∞ ≤ L. Thanks to Definition 7.1(i) we can compute Z Z T dX Z 1 + |X(t, x)| dt (t, x) dx ≤ sup log dtdx 1+R BR (0) 0 1 + |X(t, x)| BR (0) 0≤t≤T Z TZ |b(t, X(t, x))| = dxdt 0 BR (0) 1 + |X(t, x)| Z TZ |b(t, x)| φt dxdt . ≤ 0 Rd 1 + |x| Using the H¨older inequality, for every decomposition of b(t, x)/(1+|x|) as in assumption (R2) we get Z 1 + |X(t, x)| sup log dx ≤ Lk˜b1 kL1 (L1 ) + ωd Rd k˜b2 kL1 (L∞ ) . 1 + R BR (0) 0≤t≤T From this estimate we easily obtain 1 + λ −1 ˜ d Lkb1 kL1 (L1 ) + ωd Rd k˜b2 kL1 (L∞ ) , L (BR (0) \ Gλ ) ≤ log 1+R and the right hand side clearly has the properties of the function g(R, λ) stated in the proposition. 3.1. Estimate of an integral quantity and Lipschitz estimates. We start with the definition of an integral quantity which is a generalization of the quantity Ap (R, X) of Theorem 7.2. In this new setting we will need a third variable (the truncation parameter λ), hence we define Ap (R, λ, X) to be (7.14) "Z !p # p1 Z |X(t, x) − X(t, y)| sup sup − log + 1 dy dx , r BR (0)∩Gλ 0≤t≤T 0 1 and every τ ∈ [0, T ] the following estimate holds Z C ˜ x)| dx ≤ + Cλ kb − ˜bkL1 ([0,τ ]×Bλ (0)) , (7.18) 1 ∧ |X(τ, x) − X(τ, log(λ) Br (0) ˜ and on the L1 (L1 ) + L1 (L∞ ) norm of some where the constant C only depends on L, L decomposition of b and ˜b as in assumption (R2), while the constant Cλ depends on λ, ˜ and kDbkL1 ([0,τ ];Lp (B (0))) . r, L, L 3λ
3. ESTIMATES FOR MORE GENERAL VECTOR FIELDS AND COROLLARIES
119
˜ λ , relatively to X and Proof. For any given λ > 1 fixed define the sets Gλ and G ˜ X, as in (7.13). Set δ = δ(λ) = kb − ˜bkL1 ([0,τ ]×Bλ (0)) . Define Z log
g(t) = ˜λ Br (0)∩Gλ ∩G
˜ x) |X(t, x) − X(t, +1 δ
! dx .
Clearly we have g(0) = 0 and we can estimate g 0 (t) ≤
˜ x))| |b(t, X(t, x)) − ˜b(t, X(t, dx ˜ x)| + δ |X(t, x) − X(t, ˜ x)) − ˜b(t, X(t, ˜ x))| |b(t, X(t, x)) − b(t, X(t, ˜ x))| |b(t, X(t, + dx ˜ x)| + δ ˜ x)| + δ |X(t, x) − X(t, |X(t, x) − X(t,
Z ˜λ Br (0)∩Gλ ∩G
Z ≤ ˜λ Br (0)∩Gλ ∩G
˜ 1 ˜ x)) − ˜b(t, X(t, ˜ x))| + |b(t, X(t, x)) − b(t, X(t, x))| dx |b(t, X(t, ˜ x)| δ |X(t, x) − X(t,
Z ≤ ˜λ Br (0)∩Gλ ∩G
≤
1 δ
Z
˜ x)) − ˜b(t, X(t, ˜ x))| dx |b(t, X(t,
˜λ Br (0)∩Gλ ∩G
Z +cd ≤ ≤
˜λ Br (0)∩Gλ ∩G
˜ x)) dx M2λ Db(t, X(t, x)) + M2λ Db(t, X(t,
Z ˜Z L ˜ ˜ |b(t, x) − b(t, x)| dx + cd (L + L) M2λ Db(t, x) dx δ Bλ (0) Bλ (0) ˜Z L ˜ n−n/p kDb(t, ·)kLp (B (0)) . |b(t, x) − ˜b(t, x)| dx + cd,p (L + L)λ 3λ δ Bλ (0)
Integrating with respect to t between 0 and τ we obtain Z g(τ ) =
log ˜λ Br (0)∩Gλ ∩G
˜ x) |X(τ, x) − X(τ, +1 δ
! dx
˜ + cd,p (L + L)λ ˜ n−n/p kDbkL1 ([0,τ ];Lp (B (0))) = Cλ , ≤ L 3λ where the constant Cλ depends on λ but also on the other parameters relative to b and ˜b. Now fix a value η > 0 which will be specified later. We can find a measurable set ˜ λ such that L d (Br (0) ∩ Gλ ∩ G ˜ λ ) \ K < η and K ⊂ Br (0) ∩ Gλ ∩ G
log
˜ x) |X(τ, x) − X(τ, +1 δ
! ≤
Cλ η
∀x ∈ K .
120
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
Then we deduce that Z ˜ x)| dx 1 ∧ |X(τ, x) − X(τ, Br (0)
˜ λ ) + L d (Br (0) ∩ Gλ ∩ G ˜ λ) \ K + ≤ L d Br (0) \ (Gλ ∩ G
Z
˜ x)| dx |X(t, x) − X(t,
K
≤
C C + η + Cδ exp(Cλ /η) ≤ + Cλ kb − ˜bkL1 ([0,τ ]×Bλ (0)) , log(λ) log(λ)
choosing η = 1/ log(λ) in the last line.
Corollary 7.23 (Stability of the flow). Let {bh } be a sequence of vector fields satisfying assumptions (R1) and (R2), converging in L1loc ([0, T ] × Rd ) to a vector field b which satisfies assumptions (R1) and (R2). Denote by X and Xh the regular Lagrangian flows associated to b and bh respectively, and denote by L and Lh the compressibility constants of the flows. Suppose that, for some decomposition bh /(1 + |x|) = ˜bh,1 + ˜bh,2 as in assumption (R2), we have k˜bh,1 kL1 (L1 ) + k˜bh,2 kL1 (L∞ ) equi-bounded in h and that the sequence {Lh } is equi-bounded. Then the sequence {Xh } converges to X locally in measure in [0, T ] × Rd . Proof. Notice that, under the hypotheses of this corollary, the constants C h,τ and in (7.18) can be chosen uniformly with respect to τ ∈ [0, T ] and h ∈ N. Hence we find universal constant C and Cλ , depending only on the assumed equi-bounds, such that Z C h,τ 1 ∧ |X(τ, x) − Xh (τ, x)| dx ≤ + Cλh,τ kb − bh kL1 ([0,τ ]×Bλ (0)) log(λ) Br (0) C (7.19) ≤ + Cλ kb − bh kL1 ([0,T ]×Bλ (0)) . log(λ) Cλh,τ
Now fix > 0. We first choose λ big enough such that C ≤ , log(λ) 2 where C is the first constant in (7.19). Since now λ is fixed, we find N such that for every h ≥ N we have , kb − bh kL1 ([0,T ]×Bλ (0)) ≤ 2Cλ thanks to the convergence of the sequence {bh } to b in L1loc ([0, T ] × Rd ). Notice that N depends on λ and on the equi-bounds, but in turn λ only depends on and on the equi-bounds. Hence we get Z 1 ∧ |X(τ, x) − Xh (τ, x)| dx ≤ for every h ≥ N = N (). Br (0)
This means that {Xh (τ, ·)} converges to X(τ, ·) locally in measure in Rd , uniformly with respect to τ ∈ [0, T ]. In particular we get the thesis.
4. A DIRECT PROOF OF COMPACTNESS
121
Corollary 7.24 (Uniqueness of the flow). Let b be a vector field satisfying assumptions (R1) and (R2). Then the regular Lagrangian flow associated to b, if it exists, is unique. Proof. It follows immediately from Corollary 7.23.
4. A direct proof of compactness In this section we propose an alternative proof of the compactness result of Theorem 7.8, which works under an assumption of summability of the maximal function of Db. The strategy of this proof is slightly different from the previous one: we are not going to use the Lipschitz estimates of Proposition 7.4 and Remark 7.5, but instead we prove an estimate of an integral quantity which turns out to be sufficient to get compactness, via the Riesz-Fr´echet-Kolmogorov compactness criterion. We will assume the following regularity assumption on the vector field: (R3) For every λ > 0 we have Mλ Db ∈ L1 ([0, T ]; L1loc (Rd )). Notice that, by Lemma A.8, this assumption is equivalent to the condition Z TZ |Dx b(t, x)| log (2 + |Dx b(t, x)|) dxdt < ∞ for every ρ > 0. 0
Bρ (0)
This means that Dx b ∈ L1 ([0, T ]; L log Lloc (Rd )), i.e. a slightly stronger bound than Dx b ∈ L1 ([0, T ], L1loc (Rd )). We define a new integral quantity, which corresponds to the one defined in Theorem 7.2 for p = 1, but without the supremum with respect to r. For R > 0 and 0 < r < R/2 fixed we set Z Z |X(t, x) − X(t, y)| + 1 dydx . a(r, R, X) = sup − log r BR (0) 0≤t≤T Br (x) We first give a quantitative estimate for the quantity a(r, R, X), similar to the one for Ap (R, X). Theorem 7.25. Let b be a bounded vector field satisfying assumption (R3) and let X be a regular Lagrangian flow associated to b, with compressibility constant L. Then we have a(r, R, X) ≤ C R, L, kMR˜ Dx bkL1 ([0,T ];L1 (BR˜ (0))) , ˜ = 3R/2 + 2T kbk∞ . where R Proof. We start as in the proof of Theorem 7.2, obtaining inequality (7.3) (but ˜ = 3R/2 + 2T kbk∞ ). Integrating with respect to the this time it is sufficient to set R time and then with respect to x over BR (0), we obtain Z Z T a(r, R, X) ≤ cR + cd MR˜ Db(t, X(t, x)) dtdx Z
BR (0) 0 Z TZ
−
+cd BR (0)
0
Br (x)
MR˜ Db(t, X(t, y)) dydtdx .
122
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
As in the previous computations, the first integral can be estimated with
, cd L MR˜ Db L1 ([0,T ];L1 (B (0))) R+T kbk∞
but this time we cannot bound the norm of the maximal function with the norm of the derivative. To estimate the last integral we compute Z Z TZ MR˜ Db(t, X(t, y)) dydtdx cd − BR (0)
Z
0
Z
Br (x) TZ
−
= cd BR (0)
Z ≤ cd − Br (0)
Z ≤ cd − Br (0)
Z
0 T
0
Z
Br (0)
MR˜ Db(t, X(t, x + z)) dzdtdx
Z BR (0)
T
MR˜ Db(t, X(t, x + z)) dxdtdz
Z L
0
B3R/2+T kbk∞ (0)
MR˜ Db(t, w) dwdtdx
= cd LkMR˜ DbkL1 ([0,T ];L1 (B3R/2+T kbk∞ (0))) . ˜ Hence the thesis follows, by definition of R.
Next we show in Corollary 7.27 how this estimate implies compactness for the flow. We will need the following well-known criterion for strong compactness in Lp . Lemma 7.26 (Riesz-Fr´echet-Kolmogorov compactness criterion). Let F be a bounded subset of Lp (RN ) for some 1 ≤ p < ∞. Suppose that lim kf (· − h) − f kp = 0
|h|→0
uniformly in f ∈ F .
Then F is relatively compact in Lploc (RN ). Corollary 7.27 (Compactness of the flow). Let {bh } be a sequence of vector fields equi-bounded in L∞ ([0, T ] × Rd ; Rd ) and suppose that the sequence {Mλ Dbh } is equibounded in L1 ([0, T ]; L1loc (Rd )) for every λ > 0. For each h, let Xh be a regular Lagrangian flow associated to bh and let Lh be the compressibility constant associated to Xh , as in Definition 7.1(ii). Suppose that the sequence {Lh } is equi-bounded. Then the sequence {Xh } is strongly precompact in L1loc ([0, T ] × Rd ). Proof. We apply Theorem 7.25 to obtain that, under the assumptions of the corollary, the quantities a(r, R, Xh ) are uniformly bounded with respect to h. Now ˜ (with R ˜ = 3R/2 + 2T kbk∞ as in Theorem 7.25), thanks to observe that, for 0 ≤ z ≤ R the concavity of the logarithm we have ˜ log Rr + 1 z +1 ≥ z. log ˜ r R
4. A DIRECT PROOF OF COMPACTNESS
123
˜ this implies that Since |Xh (t, x) − Xh (t, y)| ≤ R Z Z |Xh (t, x) − Xd (t, y)| dydx sup − BR (0) 0≤t≤T Br (x)
˜ R
≤ log
˜ R r
+1
C R, Lh , kMR˜ Dbh kL1 ([0,T ];L1 (B ˜ (0))) ≤ g(r) , R
where the function g(r) does not depend on h and satisfies g(r) ↓ 0 for r ↓ 0. Changing the integration order this implies Z Z − |Xh (t, x) − Xh (t, x + z)| dxdz ≤ g(r) , Br (0)
BR (0)
uniformly with respect to t and h. Now notice the following elementary fact. αd > 0 with the following property: if A L d (B1 (0) \ A) ≤ αd , then A + A ⊃ B1/2 (0). could find x ∈ B1/2 (0) such that x 6∈ A + A. x 6∈ (A ∩ B1/2 (0)) + (A ∩ B1/2 (0)), so that
There exists a dimensional constant ⊂ B1 (0) is a measurable set with Indeed, if the thesis were false, we This would imply in particular that
x − A ∩ B1/2 (0) ∩ A ∩ B1/2 (0) = ∅ .
(7.20)
Now notice that there exists a dimensional constant γd such that L d B1/2 (0) ∩ (x − B1/2 (0)) ≥ γd , since we are supposing x ∈ B1/2 (0). But since L d (B1 (0) \ A) ≤ αd , we also have L d B1/2 (0) \ A ∩ B1/2 (0) ≤ αd and Ld
x − B1/2 (0) \ x − A ∩ B1/2 (0) = L d B1/2 (0) \ A ∩ B1/2 (0) ≤ αd .
But this is clearly in contradiction with (7.20) if we choose αd < γd /2. Then fix αd as above and apply the Chebyshev inequality for every h to obtain, for every 0 < r < R/2, a measurable set Kr,h ⊂ Br (0) with L d Br (0) \ Kr,h ≤ αd L d (Br (0)) and Z g(r) |Xh (t, x + z) − Xh (t, x)| dx ≤ for every z ∈ Kr,h . αd BR (0) For such a set Kr,h , thanks to the previous remark, we have that Kr,h + Kr,h ⊃ Br/2 (0). Now let v ∈ Br/2 (0) be arbitrary. For every h we can write v = z1,h + z2,h with
124
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
z1,h , z2,h ∈ Kr,h . We can estimate the increment in the spatial directions as follows: Z |Xh (t, x + v) − Xh (t, x)| dx BR/2 (0)
Z |Xh (t, x + z1,h + z2,h ) − Xh (t, x)| dx
= BR/2 (0)
Z |Xh (t, x + z1,h + z2,h ) − Xh (t, x + z1,h )| + |Xh (t, x + z1,h ) − Xh (t, x)| dx
≤ BR/2 (0)
Z
Z
≤
|Xh (t, y + z2,h ) − Xh (t, y)| dy + BR (0)
|Xh (t, x + z1,h ) − Xh (t, x)| dx ≤ BR (0)
2g(r) . αd
Now notice that, by Definition 7.1(i), for L d -a.e. x ∈ Rd we have dXh (t, x) = bh (t, Xh (t, x)) dt
for every t ∈ [0, T ].
Then we can estimate the increment in the time direction in the following way Z τ dXh |Xh (t + τ, x) − Xh (t, x)| ≤ dt (t + s, x) ds Z0 τ |bh (t + s, Xh (t + s, x))| ds ≤ τ kbh k∞ . = 0
Combining these two informations, for (t0 , t1 ) ⊂⊂ [0, T ], R > 0, v ∈ Br/2 (0) and τ > 0 sufficiently small we can estimate Z t1 Z |Xh (t + τ, x + v) − Xh (t, x)| dxdt t0
Z
t1
BR/2 (0)
Z
≤
|Xh (t + τ, x + v) − Xh (t + τ, x)| + |Xh (t + τ, x) − Xh (t, x)| dxdt t0
≤ T
BR/2 (0)
2g(r) + αd
Z
t1
Z τ kbh k∞ dxdt ≤ T
t0
BR/2 (0)
2g(r) + cd T Rd τ kbh k∞ . αd
The thesis follows applying the Riesz-Fr´echet-Kolmogorov compactness criterion (see Lemma 7.26), recalling that {bh } is uniformly bounded in L∞ ([0, T ] × Rd ). Corollary 7.28 (Existence of the flow). Let b be a bounded vector field satisfying assumption (R3) and such that [div b]− ∈ L1 ([0, T ]; L∞ (Rd )). Then there exists a regular Lagrangian flow associated to b. Proof. It is sufficient to regularize b with a positive convolution kernel in Rd and apply Corollary 7.27. It is simple to check that the regularized vector fields satisfy the equi-bounds needed for the compactness result, due to the convexity of the map z 7→ z log(2 + z) for z ≥ 0.
5. Lipexpp –REGULARITY FOR TRANSPORT EQUATIONS WITH W 1,p COEFFICIENTS
125
5. Lipexpp –regularity for transport equations with W 1,p coefficients In this section we show that solutions to transport equations with Sobolev coefficients propagate a very mild regularity property of the initial data; compare the result of Theorem 7.31 with the examples of Section 5.4. Definition 7.29 (The space Lipexpp ). We say that a function f : E ⊂⊂ Rd → Rk belongs to Lipexpp (E) if for every > 0 there exists a measurable set K ⊂ E such that (i) L d (E \ K) ≤ ; (ii) Lip(f |K ) ≤ exp C−1/p for some constant C < ∞ independent on . Moreover we denote by |f |LEp (E) the smallest constant C such that the conditions above hold. Remark 7.30. Note that: • Lipexp∞ is the space of functions which coincide with a Lipschitz function almost everywhere; • |f |LEp (E) is not homogeneus, and then it is not a norm, and can be explicitely defined as n n oo |f |LEp (E) = sup 1/p log min Lip(f |K ) : L d (E \ K) ≤ ; >0
• one can compare this definition with a similar result for Sobolev functions: if f ∈ W 1,p (E; Rk ), then for every > 0 there exists a set K ⊂ E such that L d (E \ K) ≤ and Lip(f |K ) ≤ kDf kLp (E) −1/p . Theorem 7.31. Let b be a vector field satisfying assumptions (R1) and (R2) and such that div b ∈ L1 ([0, T ]; L∞ (Rd )). Let u ¯ ∈ L∞ (Rd ) such that u ¯ ∈ Lipexpp (Ω) for d every Ω ⊂⊂ R . Let u be the solution of the Cauchy problem ( ∂t u(t, x) + b(t, x) · ∇u(t, x) = 0 (7.21) u(0, ·) = u ¯. Then for every Ω ⊂⊂ Rd we have that sup |u(t, ·)|LEp (Ω) < ∞ .
0≤t≤T
Remark 7.32. Recalling Remark 2.2 we observe that we can define u(t, ·) for every t ∈ [0, T ] in such a way that u ∈ C([0, T ], L1loc (Rd ) − w). Proof of Theorem 7.31. Let X be the regular Lagrangian flow generated by b. Then: (a) There exists a constant C > 0 such that C −1 L d (Ω) ≤ L d X(t, Ω) ≤ CL d (Ω) for every t ∈ [0, T ] and for every Ω ⊂ Rd ; therefore, for every t ∈ [0, T ], we can define Ψ(t, x) via the identity X(t, Ψ(t, x)) = Ψ(t, X(t, x)) = x for L d -a.e. x ∈ Rd ; (b) For every t we have u(t, x) = u(Ψ(t, x)) for almost every x.
126
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
Note that if for every t we consider the regular Lagrangian flow Φ(t, ·, ·) of dΦ (t, τ, x) = −b(t − τ, Φ(t, τ, x)) dτ Φ(t, 0, x) = x , then Ψ(t, x) = Φ(t, t, x). Therefore, thanks to Proposition 7.18 we conclude that sup |Ψ(t, ·)|LEp (Ω) ≤ C1 (Ω)
0≤t≤T
for every Ω ⊂⊂ Rd . Let t ∈ [0, T ], R > 0 and > 0 be given. Choose K1 ⊂ BR (0) such that • L d BR (0) \ K1 ≤ /3, • Lip(Ψ(t, ·)|K1 ) ≤ exp(|Ψ(t, ·)|LEp (BR (0)) (/3)−1/p ). ¯ > 0 such that Applying Proposition 7.16 we can find R , L d Ψ(t, BR (0)) \ BR¯ (0) ≤ 3C where C is the constant in (a). Now, select K2 ⊂ BR¯ (0) such that • L d BR¯ (0) \ K2 ≤ /3C, • Lip(¯ u|K2 ) ≤ exp(|¯ u|LEp (BC(R) (0)) (/3C)−1/p ), where again C is as in (a). Next consider K = K1 ∩ (Ψ(t, ·))−1 (K2 ) = K1 ∩ X(t, K2 ). Since BR (0) \ K ⊂ BR (0) \ K1 ∪ BR (0) \ X(t, K2 ) ⊂ BR (0) \ K1 ∪ X t, Ψ(t, BR (0)) \ BR¯ (0) ∪ X(t, BR¯ (0) \ K2 ) , we have L d BR (0) \ K ≤ L d BR (0) \ K1 + L d X t, Ψ(t, BR (0)) \ BR¯ (0) + L d X(t, BR¯ (0) \ K2 ) ≤ . Given x, y ∈ K we have Ψ(t, x), Ψ(t, y) ∈ K2 and hence we can estimate ¯(Ψ(t, y))| ≤ Lip(¯ u|K2 )|Ψ(t, x) − Ψ(t, y)| |u(t, x) − u(t, y)| = |u(Ψ(t, x)) − u ≤ Lip(¯ u|K2 )Lip(Ψ(t, ·)|K1 )|x − y| nh i o = |x − y| exp (3C)1/p |¯ u|LEp (BR¯ (0)) + 31/p |Ψ(t, ·)|LEp (BR (0)) −1/p . Therefore 1/p log(Lip(u(t, ·)|K )) is bounded by a constant independent of and t (but which depends on R). Taking the supremum over t and , we conclude that sup |u(t, ·)|LEp (BR (0)) ≤ C(R) ,
0≤t≤T
and this concludes the proof.
6. AN APPLICATION TO A CONJECTURE ON MIXING FLOWS
127
6. An application to a conjecture on mixing flows In [47] the author considers a problem on mixing vector fields on the two-dimensional torus K = R2 /Z2 . In this section, we are going to show that the Lipschitz estimate of Proposition 7.18 gives an answer to this problem, although in the Lp setting (p > 1) instead of the L1 setting considered in [47]. Fix coordinates x = (x1 , x2 ) ∈ [0, 1[×[0, 1[ on K and consider the set A = (x1 , x2 ) : 0 ≤ x2 ≤ 1/2 ⊂ K . If b : [0, 1] × K → R2 is a smooth time-dependent vector field, we denote as usual by X(t, x) the flow of b and by Φ : K → K the value of the flow at time t = 1. We assume that the flow is nearly incompressibile, so that for some κ0 > 0 we have 1 (7.22) κ0 L 2 (Ω) ≤ L 2 X(t, Ω) ≤ 0 L 2 (Ω) κ for all Ω ⊂ K and all t ∈ [0, 1]. For a fixed 0 < κ < 1/2, we say that Φ mixes the set A up to scale if for every ball B (x) we have κL 2 (B (x)) ≤ L 2 B (x) ∩ Φ(A) ≤ (1 − κ)L 2 (B (x)) . Then in [47] the following conjecture is proposed: Conjecture 7.33 (Bressan’s mixing conjecture). Under these assumptions, there exists a constant C depending only on κ and κ0 such that, if Φ mixes the set A up to scale , then Z 1Z |Dx b| dxdt ≥ C| log | for every 0 < < 1/4. 0
K
In this section we show the following result: Theorem 7.34. Let p > 1. Under the previous assumptions, there exists a constant C depending only on κ, κ0 and p such that, if Φ mixes the set A up to scale , then Z 1 kDx bkLp (K) dt ≥ C| log | for every 0 < < 1/4. 0
Proof. We set M = kDx bkL1 ([0,1];Lp (K)) and A0 = K \ A. Applying Proposition 7.18, and noticing that the flow is bounded since we are on the torus, for every constant η > 0 we can find a set B with L 2 (B) ≤ η such that (7.23) Lip Φ−1 |K\B ≤ exp(βM ) , where the constant β depends only on κ0 , η and p. Since Φ mixes the set A up to scale , for every x ∈ A we have (7.24) L 2 B (Φ(x)) ∩ Φ(A0 ) ≥ κL 2 B (Φ(x)) . We define A˜ = x ∈ A : B (Φ(x)) ∩ Φ(A0 ) \ B = ∅ . From this definition and from (7.24) we get that for every x ∈ A˜ (7.25) L 2 B (Φ(x)) ∩ B ≥ κL 2 B (Φ(x)) .
128
7. A PRIORI ESTIMATES FOR REGULAR LAGRANGIAN FLOWS
From (7.25) and the Besicovitch covering theorem we deduce that for an absolute constant c we have ˜ ≤ c L 2 (B) ≤ cη . L 2 Φ(A) κ κ From the compressibility condition (7.22) we deduce ˜ ≤ cη . L 2 (A) κκ0 Since, using again (7.22), we know that L 2 (B) η ≤ 0, L 2 Φ−1 (B) ≤ 0 κ κ 0 we can choose η > 0, depending on κ and κ only, in such a way that 1 L 2 A˜ + L 2 Φ−1 (B) ≤ . 6 −1 ˜ This implies the existence of a point x ¯ ∈ A \ A ∪ Φ (B) with dist (¯ x, A0 ) ≥ 1/6. Let ˜ we can find a point z¯ ∈ B (¯ y¯ = Φ(¯ x). Since x ¯ 6∈ A, y ) ∩ Φ(A0 ) \ B . Clearly we have −1 0 |¯ y − z¯| ≤ and (since Φ (z) ∈ A ) we also have |¯ x − Φ−1 (¯ z )| ≥ 1/6. Since y¯, z¯ 6∈ B, we can apply (7.23) to deduce 1 ≤ Lip Φ−1 |K\B ≤ exp(βM ) , 6 where now β depends only on κ, κ0 and p, since η has been fixed. This implies that 1 1 . M = kDx bkL1 ([0,1];Lp (K)) ≥ log β 6 Hence we can find 0 > 0 such that 1 M≥ | log | for every 0 < < 0 . 2β We are now going to show the thesis for every 0 < < 1/4. Indeed, suppose that the thesis is false. Then, we could find a sequence {bh } of vector fields and a sequence {h } with 0 < h < 1/4 in such a way that 1 kDx bh kL1 ([0,1];Lp (K)) ≤ | log h | h and the corresponding map Φh mixes the set A up to scale h . This implies that 1 1 kDx bh kL1 ([0,1];Lp (K)) ≤ | log h | ≤ | log 0 | −→ 0 as h → ∞. h h Up to an extraction of a subsequence, we can assume that h → ¯ and that Φh → Φ strongly in L1 (K). For this, we apply the compactness result in Theorem 7.20, noticing that (7.22) gives a uniform control on the compressibility constants of the flows and that we do not need any assumption on the growth of the vector fields, since we are on the torus and then the flow is automatically uniformly bounded. Now, notice that the mixing property is stable with respect to strong convergence: this means that Φ has to mix up to scale ¯ ≤ 1/4. But since kDx bh kL1 ([0,1];Lp (K)) → 0, we deduce that Φ is indeed a translation on K, hence it cannot mix the set A up to a scale which is smaller than 1/4. From this contradiction we get the thesis.
6. AN APPLICATION TO A CONJECTURE ON MIXING FLOWS
129
Remark 7.35. We notice that the constant 1/4 in Theorem 7.34 depends on the shape of the set A: this bound comes from the fact that a translation does not mix up to a scale < 1/4. Our proof can be easily extended to the case of a measurable set A with any shape, giving a different upper bound for the values of such that the result is true.
Appendix: Background material and technical results In this appendix we collect some results on measure theory, Lipschitz functions, BV functions (for which good general references are [17], [18], [90], [91] and [95]) and on the theory of maximal functions (for which we refer to [119]). 1. Measure theory Let X be a separable metric space. We denote by B(X) the family of the Borel subsets of X, by P(X) the family of the Borel probability measures on X, by M(X) the family of the locally finite Borel measures on X and by M+ (X) the subset of M(X) consisting of all nonnegative locally finite Borel measures on X. In a similar fashion we can consider vector-valued or matrix-valued measures on X. The total variation of a measure µ ∈ M(X) is denoted by |µ|. In the case when |µ|(X) < +∞ we say that µ has finite mass and we set kµkM(X) = |µ|(X). The support of a measure µ ∈ M(X) is the closed set defined by spt µ = {x ∈ X : |µ|(U ) > 0 for each open neighborhood U of x } . A measure µ ∈ M(X) is concentrated on a Borel set E ⊂ X if |µ|(X \ E) = 0. If µ ∈ M(X) and ν ∈ M+ (X) we say that µ is absolutely continuous with respect to ν (and we write µ ν) if |µ|(E) = 0 for every Borel set E ⊂ X such that ν(E) = 0. We say that two measures µ and ν ∈ M(X) are mutually singular (and we write µ ⊥ ν) if they are concentrated on disjoint Borel sets. If µ ∈ M(X) and E ⊂ X is a Borel set, the restriction of µ to E is the measure µ E ∈ M(X) defined by (µ E)(A) = µ(A∩E) for every Borel set A ⊂ X. If f : X → Y is a Borel map between two separable metric spaces X and Y and µ ∈ M(X) we denote by f# µ ∈ M(Y ) the push-forward of the measure µ, defined by (A.1)
(f# µ)(E) = µ(f −1 (E))
for every Borel set E ⊂ Y .
Ld
We denote by the Lebesgue measure on Rd and with H k the k-dimensional Hausdorff measure on Rd . For any subset A of Rd we define the characteristic function 1A as 1 if x ∈ A 1A (x) = 0 if x 6∈ A. A family F ⊂ M+ (X) is bounded if there exists a constant C with kµkM(X) ≤ C for each µ ∈ F . We say that a bounded sequence {µh } ⊂ M+ (X) is narrowly convergent to µ ∈ M+ (X) as h → ∞ if Z Z lim f (x) dµh (x) = f (x) dµ(x) h→∞ X
X 131
132
APPENDIX: BACKGROUND MATERIAL AND TECHNICAL RESULTS
for every f ∈ Cb (X), the space of continuous and bounded real functions defined on X. We say that a bounded family F ⊂ M+ (X) is tight if for every > 0 there exists a compact set K ⊂ X such that µ(X \ K ) ≤ for every µ ∈ F . The following theorem characterizes the relatively compact subsets with respect to the narrow topology. Theorem A.1 (Prokhorov). Assume that X is a complete metric space. Then a bounded family F ⊂ M+ (X) is relatively compact with respect to the narrow convergence if and only if it is tight. Moreover, a necessary and sufficient condition for tightness is the existence of a R coercive functional Ψ : X → [0, +∞] such that X Ψ dµ ≤ 1 for every µ ∈ F . It is also simple to check that a bounded family F ⊂ M+ (X × Y ) is tight if and only if the families of the marginals πX # F ⊂ M+ (X) and πY # F ⊂ M+ (Y ) are tight. We also recall a particular case of the disintegration theorem. Let µ ∈ M+ (X × Y ), define ν = (πX )# µ and assume that ν ∈ M+ (X). Then there exists a Borel family {µx }x∈X ⊂ M+ (Y ), which is uniquely determined ν-a.e., such that Z µ= µx dν(x) . X
Fix a measure µ ∈ M+ (X). We say that a sequence of functions {fh } defined on X and with values in Rk converges in µ-measure to a function f if lim µ {x ∈ X : |fh (x) − f (x)| > δ} = 0 for every δ > 0. h→∞
We say that a Borel set Σ ⊂ Rd is H k -rectifiable if there exist countably many Lipschitz functions fi : Rk → Rd such that [ H k Σ \ fi Rk = 0 . i
We finally recall the coarea formula for Lipschitz functions (see for instance Section 2.12 of [17]): for every Lipschitz map f : Rd → Rd−k and every positive Borel function ϕ : Rd → [0, +∞] there holds Z Z Z d k (A.2) ϕJ dL = ϕ dH dL d−k (h) , Rd
Rd−k
Eh
t ∇f )]1/2
where J = [det(∇f · is the Jacobian of f and Eh = {x ∈ Rd : f (x) = h}, for h ∈ Rd−k , are the level sets of the function f . If φ ∈ L1 (Rd ) from (A.2) we also deduce Z Z Z φ k d dH dL d−k (h) . (A.3) φ dL = Rd Rd−k Eh J 2. Lipschitz functions, extension theorems and approximate differentiability A map f : Ω ⊂ Rd → Rk is Lipschitz if there exists a constant L such that (A.4)
|f (x) − f (y)| ≤ L|x − y|
for every x, y ∈ Ω.
The minimal constant L such that (A.4) holds is called the Lipschitz constant of the function f and is denoted by Lip(f ).
3. FUNCTIONS WITH BOUNDED VARIATION
133
Theorem A.2 (Rademacher). A Lipschitz function f : Rd → Rk is differentiable at L d -a.e. x ∈ Rd . It is simple to check that every real-valued Lipschitz function f : Ω ⊂ Rd → R can be extended to a function f˜ : Rd → R with Lip(f˜) = Lip(f ) (by extension we mean that f˜|Ω = f ). Indeed, it is sufficient to set f˜ = f + or f˜ = f − , where f ± (x) = inf f (y) ± L|x − y| : y ∈ Ω and L = Lip(f ). Let us notice that f + and f − are respectively the biggest and the smallest extensions. The same result holds also if Ω is a metric space. Moreover, arguing componentwise, we deduce that every Lipschitz√function f : Ω ⊂ Rd → Rk can be extended to a function f˜ : Rd → Rk with Lip(f˜) = k Lip(f ). In fact, for subsets of Rd , also the following stronger result holds: Theorem A.3 (Kirszbraun). Let f : Ω ⊂ Rd → Rk be a Lipschitz function. Then there exists an extension f˜ : Rd → Rk with Lip(f˜) = Lip(f ). Its proof is considerably more difficult and we mention it for completeness, although it is not strictly necessary in our presentation. We say that a Borel map f : Rd → Rk is approximately differentiable at x ∈ Rd if there exists a linear map L : Rd → Rk such that the difference quotients f (x + y) − f (x) locally converge in measure as ↓ 0 to Ly. This is clearly a local property. It is possible to show that, if f |K is a Lipschitz map for some set K ⊂ Rd , then f is approximately differentiable at almost every point of K. In the following theorem we show a kind of converse of this statement: an approximately differentiable map can be approximated, in the Lusin sense, with Lipschitz maps. y 7→
Theorem A.4. Let f : Ω → Rk . Assume that there exists a sequence of Borel sets Ah ⊂ Ω such that L d Ω \ ∪h Ah = 0 and f |Ah is Lipschitz for any h. Then f is approximately differentiable at L d -a.e. x ∈ Ω. Conversely, if f is approximately differentiable at all points of Ω0 ⊂ Ω, we can write Ω0 as a countable union of sets Ah such that f |Ah is Lipschitz for any h (up to a redefinition of f in an L d -negligible set). 3. Functions with bounded variation We say that a function f : Ω ⊂ Rd → R has bounded variation, and we write f ∈ BV (Ω), if f ∈ L1 (Ω) and the distribution derivative Df is a vector-valued finite measure in Ω. We also introduce the space BVloc (Ω) of functions with locally bounded ˜ for every variation as the class of those f : Ω ⊂ Rd → R such that f ∈ BV (Ω) ˜ ⊂⊂ Ω. The spaces BV (Ω; Rk ) and BVloc (Ω; Rk ) are defined by requiring BV or Ω BVloc regularity on each component. The space BV (Ω) is a Banach space with the norm kf kBV (Ω) = kf kL1 (Ω) + kDf kM(Ω) .
134
APPENDIX: BACKGROUND MATERIAL AND TECHNICAL RESULTS
When f ∈ L1 (Ω; Rk ) we can consider the set of points of approximate discontinuity Sf defined by ( ) Z Ω \ Sf = x ∈ Ω : ∃z ∈ Rk s.t. − |f (y) − z| dy → 0 as r → 0 . Br (x)
Every x ∈ Ω\Sf is called a point of approximate continuity. The value z ∈ Rk appearing in the definition, when it exists, is unique and is denoted by f˜(x). It is simple to check that Sf is a Borel set and that the equality f = f˜ holds L d -a.e. in Ω \ Sf . We also introduce the set of points of approximate jump Jf ⊂ Sf : ) ( Z |f (y) − z ± | dy → 0 as r → 0 , Jf = x ∈ Sf : ∃ν ∈ Sd−1 , ∃z ± ∈ Rk s.t. − ± Bν,r (x)
where we have defined the half-balls ± Bν,r (x) = {y ∈ Br (x) : (y − x) · ν ≷ 0} .
The triple (ν, z − , z + ), when it exists, is unique up to a permutation of z − and z + and a change of sign of ν, and is denoted by (ν(x), f − (x), f + (x)). Notice that f − (x) 6= f + (x) when they exist, since x ∈ Sf . It is easy to check that Jf is a Borel set and that f ± and ν can be chosen to be Borel functions in their domain of definition. For f ∈ BV (Ω; Rk ), with Ω ⊂ Rd , we can use the Lebesgue decomposition theorem to obtain Df = Da f + Dc f , with Da f L d and Ds ⊥ L d . The measure Da f is called the absolutely continuous part of the derivative and the measure Ds f is called the singular part of the derivative. The singular part can be decomposed as Ds f = Dj f + Dc f = Ds f
Jf + Ds f (Ω \ Jf ) ,
where Dj f is the jump part of the derivative and Dc f is the Cantor part of the derivative. The subspace SBV (Ω; Rk ) ⊂ BV (Ω; Rk ) of special function with bounded variation consists of those f ∈ BV (Ω; Rk ) with Dc f = 0. The following important result relative to the structure of BV functions holds. For the notion of rectifiable set we refer to Appendix 1. Theorem A.5 (Structure of BV functions). Let f ∈ BV (Ω; Rk ). Then the approximate jump set is countably H d−1 -rectifiable, H d−1 (Sf \ Jf ) = 0 and (A.5) Dj f = f + − f − ⊗ ν H d−1 Jf . The following deep result by Alberti (see [3] and the recent simplified presentation of [77]) says that the rank-one structure of the jump part of the derivative expressed by (A.5) is also shared by the Cantor part. Theorem A.6 (Alberti’s Rank-one theorem). Let f ∈ BV (Ω; Rk ). Then there exist two Borel functions ξ : Ω → Sd−1 and η : Ω → Sk−1 such that Ds f = ξ ⊗ η|Ds f | .
4. MAXIMAL FUNCTIONS
135
In a similar fashion we can define the space of vector fields with bounded deformation in an open set Ω ⊂ Rd by requiring that the simmetric part of the distributional derivative Ef = 21 Df + t Df is a measure: n o BD(Ω; Rd ) = f ∈ L1 (Ω; Rd ) : Ef is a matrix valued finite measure in Ω . We refer to [122] and [12] for a presentation of the main properties of this class of vector fields. As in the BV case we can decompose Ef = E a f + E j f + E c f and define the space of special vector fields with bounded deformation as n o SBD(Ω; Rd ) = f ∈ BD(Ω; Rd ) : E c f = 0 . 4. Maximal functions We recall here the definition of the local maximal function of a locally finite measure and of a locally summable function and we state some well-known properties. Definition A.7 (Local maximal function). Let µ be a (vector-valued) locally finite measure. For every λ > 0, we define the local maximal function of µ as Z |µ|(Br (x)) d|µ|(y) x ∈ Rd . = sup − Mλ µ(x) = sup d (B (x)) L r 0 t}) ≤ |f (y)| dy , t Bρ+λ (0) for every t > 0. Lemma A.9. If u ∈ BV (Rd ) then there exists an L d -negligible set N ⊂ Rd such that |u(x) − u(y)| ≤ cd |x − y| (Mλ Du(x) + Mλ Du(y)) for x, y ∈ Rd \ N with |x − y| ≤ λ.
136
APPENDIX: BACKGROUND MATERIAL AND TECHNICAL RESULTS
We also recall the Chebyshev inequality: for every t > 0 we have Z L d ({|f | > t})1/q 1 d |f (x)| dx ≤ L ({|f | > t}) ≤ kf kLp (Ω) , t {|f |>t} t where (A.7)
1 p
+
1 q
= 1, and this implies L d ({|f | > t})1/p ≤
kf kLp (Ω) . t
Bibliography [1] R. A. Adams: Sobolev spaces. Pure and Applied Mathematics, 65 (1975), Academic Press, New York–London. [2] M. Aizenman: On vector fields as generators of flows: a counterexample to Nelson’s conjecture. Ann. Math., 107 (1978), 287–296. [3] G. Alberti: Rank-one properties for derivatives of functions with bounded variation. Proc. Roy. Soc. Edinburgh Sect. A, 123 (1993), 239–274. [4] G. Alberti & L. Ambrosio: A geometric approach to monotone functions in Rn . Math. Z., 230 (1999), 259–316. [5] G. Alberti, S. Bianchini & G. Crippa: Work in preparation. ¨ller: A new approach to variational problems with multiple scales. Comm. [6] G. Alberti & S. Mu Pure Appl. Math., 54 (2001), 761–825. [7] F. J. Almgren: The theory of varifolds – A variational calculus in the large. Princeton University Press, 1972. [8] L. Ambrosio: Transport equation and Cauchy problem for BV vector fields. Invent. Math., 158 (2004), 227–260. [9] L. Ambrosio: Lecture notes on transport equation and Cauchy problem for BV vector fields and applications. Preprint, 2004 (available at http://cvgmt.sns.it). [10] L. Ambrosio: Lecture notes on transport equation and Cauchy problem for non-smooth vector fields and applications. Preprint, 2005 (available at http://cvgmt.sns.it). [11] L. Ambrosio, F. Bouchut & C. De Lellis: Well-posedness for a class of hyperbolic systems of conservation laws in several space dimensions. Comm. PDE, 29 (2004), 1635–1651. [12] L. Ambrosio, A. Coscia & G. Dal Maso: Fine properties of functions with bounded deformation. Arch. Rational Mech. Anal., 139 (1997), 201–238. [13] L. Ambrosio & G. Crippa: Existence, uniqueness, stability and differentiability properties of the flow associated to weakly differentiable vector fields. To appear in the series UMI Lecture Notes (available at http://cvgmt.sns.it). [14] L. Ambrosio, G. Crippa & S. Maniglia: Traces and fine properties of a BD class of vector fields and applications. Ann. Sci. Toulouse, XIV (4) (2005), 527–561. [15] L. Ambrosio & C. De Lellis: Existence of solutions for a class of hyperbolic systems of conservation laws in several space dimensions. International Mathematical Research Notices, 41 (2003), 2205–2220. ´ : On the chain-rule for the divergence of BV like vector [16] L. Ambrosio, C. De Lellis & J. Maly fields: applications, partial results, open problems. To appear in the forthcoming book by the AMS series in contemporary mathematics “Perspectives in Nonlinear Partial Differential Equations: in honor of Haim Brezis” (available at http://cvgmt.sns.it). [17] L. Ambrosio, N. Fusco & D. Pallara: Functions of bounded variation and free discontinuity problems. Oxford Mathematical Monographs, 2000. ´: Gradient flows in metric spaces and in the Wasserstein [18] L. Ambrosio, N. Gigli & G. Savare space of probability measures. Lectures in Mathematics, ETH Zurich, Birkh¨ auser, 2005. [19] L. Ambrosio, M. Lecumberry & S. Maniglia: Lipschitz regularity and approximate differentiability of the DiPerna–Lions flow. Rend. Sem. Mat. Univ. Padova, 114 (2005), 29–50.
137
138
BIBLIOGRAPHY
´: Stability of flows associated to gradient vector fields and [20] L. Ambrosio, S. Lisini & G. Savare convergence of iterated transport maps. Manuscripta Math., 121 (2006), 1–50. ´: Very weak notions of differentiability. Proceedings of the Royal Society [21] L. Ambrosio & J. Maly of Edinburgh, 137A (2007), 447–455. [22] L. Ambrosio, P. Tilli & L. Zambotti: Introduzione alla Teoria della Misura ed alla Probabilit` a. Lecture notes of a course given at the Scuola Normale Superiore, unpublished. [23] G. Anzellotti: Pairings between measures and bounded functions and compensated compactness. Ann. Mat. Pura App., 135 (1983), 293–318. [24] G. Anzellotti: The Euler equation for functionals with linear growth. Trans. Amer. Mat. Soc., 290 (1985), 483–501. [25] G. Anzellotti: Traces of bounded vectorfields and the divergence theorem. Unpublished preprint, 1983. [26] H. Bahouri & J.-Y. Chemin: Equations de transport relatives ` a des champs de vecteurs nonlipschitziens et m´ecanique des fluides. Arch. Rational Mech. Anal., 127 (1994), 159–181. [27] E. J. Balder: New fundamentals of Young measure convergence. CRC Res. Notes in Math. 411, 2001. [28] J. Ball & R. James: Fine phase mixtures as minimizers of energy. Arch. Rational Mech. Anal., 100 (1987), 13–52. [29] V. Bangert: Minimal measures and minimizing closed normal one-currents. Geom. funct. anal., 9 (1999), 413–427. [30] A. Beck: Uniqueness of flow solutions of differential equations. Recent advances in topological dynamics. Lecture Notes in Math., Vol. 318, Springer, Berlin, 1973. [31] J.-D. Benamou & Y. Brenier: Weak solutions for the semigeostrophic equation formulated as a couples Monge-Ampere transport problem. SIAM J. Appl. Math., 58 (1998), 1450–1461. [32] P. Bernard & B. Buffoni: Optimal mass transportation and Mather theory. J. Eur. Math. Soc. (JEMS), 9 (2007), 85–121. [33] O. Besson & J. Pousin: Solutions for linear conservation laws with velocity fields in L∞ . Accepted by Arch. Rational Mech. Anal. [34] S. Bianchini & A. Bressan: Vanishing viscosity solutions of nonlinear hyperbolic systems. Ann. Math., 161 (2005), 223–342. [35] V. Bogachev & E. M. Wolf: Absolutely continuous flows generated by Sobolev class vector fields in finite and infinite dimensions. J. Funct. Anal., 167 (1999), 1–68. [36] F. Bouchut: Renormalized solutions to the Vlasov equation with coefficients of bounded variation. Arch. Rational Mech. Anal., 157 (2001), 75–90. [37] F. Bouchut & G. Crippa: Uniqueness, Renormalization, and Smooth Approximations for Linear Transport Equations. SIAM J. Math. Anal., 38 (2006), 1316–1328. [38] F. Bouchut & L. Desvillettes: On two-dimensional Hamiltonian transport equations with continuous coefficients. Differential Integral Equations, 14 (2001), 1015–1024. [39] F. Bouchut, F. Golse & M. Pulvirenti: Kinetic equations and asymptotic theory. Series in Appl. Math., Gauthiers-Villars, 2000. [40] F. Bouchut & F. James: One-dimensional transport equation with discontinuous coefficients. Nonlinear Analysis, 32 (1998), 891–933. [41] F. Bouchut, F. James & S. Mancini: Uniqueness and weak stability for multi-dimensional transport equations with one-sided Lipschitz coefficients. Annali Scuola Normale Superiore, Ser. 5, 4 (2005), 1–25. [42] Y. Brenier: The least action principle and the related concept of generalized flows for incompressible perfect fluids. J. Amer. Mat. Soc., 2 (1989), 225–255. [43] Y. Brenier: The dual least action problem for an ideal, incompressible fluid. Arch. Rational Mech. Anal., 122 (1993), 323–351. [44] Y. Brenier: A homogenized model for vortex sheets. Arch. Rational Mech. Anal., 138 (1997), 319–353. [45] Y. Brenier: Minimal geodesics on groups of volume-preserving maps and generalized solutions of the Euler equations. Comm. Pure Appl. Math., 52 (1999), 411–452.
BIBLIOGRAPHY
139
[46] A. Bressan: Hyperbolic systems of conservation laws. The one-dimensional Cauchy problem. Oxford Lecture Series in Mathematics and its Applications, 20. Oxford University Press, Oxford, 2000. [47] A. Bressan: A lemma and a conjecture on the cost of rearrangements. Rend. Sem. Mat. Univ. Padova, 110 (2003), 97–102. [48] A. Bressan: An ill posed Cauchy problem for a hyperbolic system in two space dimensions. Rend. Sem. Mat. Univ. Padova, 110 (2003), 103–117. [49] H. Brezis: Analyse fonctionnelle. Th´eorie et applications. Masson, Paris, 1983. [50] I. Capuzzo Dolcetta & B. Perthame: On some analogy between different approaches to first order PDE’s with nonsmooth coefficients. Adv. Math. Sci Appl., 6 (1996), 689–703. [51] A. Cellina: On uniqueness almost everywhere for monotonic differential inclusions. Nonlinear Analysis, TMA, 25 (1995), 899–903. [52] A. Cellina & M. Vornicescu: On gradient flows. Journal of Differential Equations, 145 (1998), 489–501. [53] J.-Y. Chemin: Perfect incompressible fluids. Oxford University Press, 1998. [54] J.-Y. Chemin & N. Lerner: Flot de champs de vecteurs non lipschitziens et ´equations de Navier– Stokes. Journal of Differential Equations, 121 (1995), 314–328. [55] G.-Q. Chen & H. Frid: Divergence-measure fields and conservation laws. Arch. Rational Mech. Anal., 147 (1999), 89–118. [56] G.-Q. Chen & H. Frid: Extended divergence-measure fields and the Euler equation of gas dynamics. Comm. Math. Phys., 236 (2003), 251–280. [57] F. Colombini, G. Crippa & J. Rauch: A note on two-dimensional transport with bounded divergence. Comm. PDE, 31 (2006), 1109–1115. ´ [58] F. Colombini & N. Lerner: Sur les champs de vecteurs peu r´eguliers. S´eminaire: Equations aux ´ D´eriv´ees Partielles, Exp. No. XIV, Ecole Polytech., Palaiseau, 2001. [59] F. Colombini & N. Lerner: Uniqueness of continuous solutions for BV vector fields. Duke Math. J., 111 (2002), 357–384. [60] F. Colombini & N. Lerner: Uniqueness of L∞ solutions for a class of conormal BV vector fields. Contemp. Math. 368 (2005), 133–156. [61] F. Colombini, T. Luo & J. Rauch: Uniqueness and nonuniqueness for nonsmooth divergence´ ´ free transport. S´eminaire: Equations aux D´eriv´ees Partielles, Exp. No. XXII, Ecole Polytech., Palaiseau, 2003. [62] F. Colombini, T. Luo & J. Rauch: Nearly Lipschitzean divergence-free transport propagates neither continuity nor BV regularity. Commun. Math. Sci., 2 (2004), 207–212. [63] F. Colombini & J. Rauch: Uniqueness in the Cauchy Problem for Transport in R2 and R1+2 . J. Differential Equations, 211 (2005), 162–167. [64] G. Crippa: Equazione del trasporto e problema di Cauchy per campi vettoriali debolmente differenziabili. Tesi di Laurea, Universit` a di Pisa, 2006 (available at http://cvgmt.sns.it). [65] G. Crippa & C. De Lellis: Oscillatory solutions to transport equations. Indiana Univ. Math. J., 55 (2006), 1–13. [66] G. Crippa & C. De Lellis: Estimates and regularity results for the DiPerna–Lions flow. Preprint, 2006 (available at http://cvgmt.sns.it). J. Reine Angew. Math., in press. [67] G. Crippa & C. De Lellis: Regularity and compactness for the DiPerna–Lions flow. Preprint, 2006 (available at http://cvgmt.sns.it). To be published in the Proceedings of HYP2006. ´ [68] A. B. Cruzeiro: Equations diff´erentielles ordinaires: non explosion et mesures quasi-invariantes. J. Funct. Anal., 54 (1983), 193–205. ´ [69] A. B. Cruzeiro: Equations diff´erentielles sur l’espace de Wiener et formules de Cameron-Martin non lin´eaires. J. Funct. Anal., 54 (1983), 206–227. [70] A. B. Cruzeiro: Unicit´e de solutions d’´equations diff´erentielles sur l’espace de Wiener. J. Funct. Anal., 58 (1984), 335–347. [71] M. Cullen: On the accuracy of the semi-geostrophic approximation. Quart. J. Roy. Metereol. Soc., 126 (2000), 1099–1115.
140
BIBLIOGRAPHY
[72] M. Cullen & M. Feldman: Lagrangian solutions of semigeostrophic equations in physical space. SIAM J. Math. Anal., 37 (2006), 1371–1395. [73] M. Cullen & W. Gangbo: A variational approach for the two-dimensional semi-geostrophic shallow water equations. Arch. Rational Mech. Anal., 156 (2001), 241–273. [74] C. Dafermos: Hyperbolic conservation laws in continuum physics. Springer Verlag, 2000. [75] C. De Lellis: Blow-up of the BV norm in the multi-dimensional Keyfitz and Kranzer system. Duke Math. J., 127 (2004), 313–339. [76] C. De Lellis: Notes on hyperbolic systems of conservation laws and transport equations. Handbook of Differential Equations: Evolutionary Equations, vol. III. Edited by C. M. Dafermos and E. Feireisl. Elsevier/North-Holland, Amsterdam, 2006. [77] C. De Lellis: A note on Alberti’s Rank-one Theorem. To appear in the series UMI Lecture Notes (available at http://cvgmt.sns.it). [78] C. De Lellis: Ordinary differential equations with rough coefficients and the renormalization theorem of Ambrosio (d’apr`es Ambrosio, DiPerna, Lions). S´eminaire Bourbaki, vol. 2006/2007, n. 972. [79] C. De Lellis & M. Westdickenberg: On the optimality of velocity averaging lemmas. Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 20 (2003), 1075–1085. [80] L. De Pascale, M. S. Gelli & L. Granieri: Minimal measures, one-dimensional currents and the Monge-Kantorovich problem. Calc. Var. Partial Differential Equations, 27 (2006), 1–23. [81] N. Depauw: Non unicit´e des solutions born´ees pour un champ de vecteurs BV en dehors d’un hyperplan. C.R. Math. Sci. Acad. Paris, 337 (2003), 249–252. [82] R. J. DiPerna: Measure-valued solutions to conservation laws. Arch. Rational Mech. Anal., 88 (1985), 223–270. [83] R. J. DiPerna & P.-L. Lions: Ordinary differential equations, transport theory and Sobolev spaces. Invent. Math., 98 (1989), 511–547. [84] R. J. DiPerna & P.-L. Lions: On the Cauchy problem for the Boltzmann equation: global existence and weak stability. Ann. of Math., 130 (1989), 312–366. [85] R. Engelking: General Topology. Revised and completed edition. Sigma series in pure mathematics 6. Heldermann Verlag, Berlin (1989). [86] L. C. Evans: Partial Differential Equations and Monge–Kantorovich Mass Transfer. Current Developments in Mathematics (1997), 65–126. [87] L. C. Evans: Partial Differential Equations. Graduate studies in Mathematics, 19 (1998), American Mathematical Society. [88] L. C. Evans & W. Gangbo: Differential equations methods for the Monge-Kantorovich mass transfer problem. Mem. Amer. Math. Soc. 137 (1999), no. 653. [89] L. C. Evans, W. Gangbo & O. Savin: Diffeomorphisms and nonlinear heat flows. SIAM J. Math. Anal., 37 (2005), 737–751. [90] L. C. Evans & R. F. Gariepy: Lecture notes on measure theory and fine properties of functions. CRC Press, 1992. [91] H. Federer: Geometric measure theory. Springer, 1969. [92] A. Figalli: Existence and uniqueness for SDEs with rough or degenerate coefficients. Preprint, 2007. Accepted by J. Funct. Anal. [93] M. Giaquinta & G. Modica: Analisi matematica III: Strutture lineari e metriche, continuit` a. Pitagora Editrice, Bologna. 2000. [94] D. Gilbarg & N. S. Trudinger: Elliptic partial differential equations of second order. Classics in Mathematics. Springer-Verlag, Berlin, 2001. [95] E. Giusti: Minimal surfaces and functions of bounded variation. Monographs in Mathematics 80, Birkh¨ auser Verlag, 1984. [96] J. Glimm: Solutions in the large for nonlinear hyperbolic systems of equations. Comm. Pure Appl. Math., 18 (1965), 697–715. [97] P. Hartman: Ordinary differential equations. Classics in Applied Mathematics 38 (2002), Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA.
BIBLIOGRAPHY
141
[98] M. Hauray: On Liouville transport equation with potential in BVloc . Comm. PDE, 29 (2004), 207–217. [99] M. Hauray: On two-dimensional Hamiltonian transport equations with Lploc coefficients. Ann. IHP Nonlinear Anal. Non Lin´eaire, 20 (2003), 625–644. [100] M. Hauray, C. Le Bris & P.-L. Lions: Deux remarques sur les flots g´en´eralis´es d’´equations diff´erentielles ordinaires. Preprint, 2007. [101] B. L. Keyfitz & H. C. Kranzer: A system of nonstrictly hyperbolic conservation laws arising in elasticity theory. Arch. Rational Mech. Anal., 72 (1980), 219–241. [102] S. N. Kruˇ zkov: First order quasilinear equations with several independent variables. Mat. Sb. (N.S.), 81 (1970), 228–255. [103] C. Le Bris & P.-L. Lions: Renormalized solutions of some transport equations with partially W 1,1 velocities and applications. Annali di Matematica, 183 (2003), 97–130. [104] N. Lerner: Transport equations with partially BV velocities. Ann. Sc. Norm. Super. Pisa Cl. Sci., 3 (2004), 681–703. ´ ´ [105] N. Lerner: Equations de transport dont les vitesses sont partiellement BV . S´eminaire: Equations ´ aux D´eriv´ees Partielles, Exp. No. X, Ecole Polytech., Palaiseau, 2004. [106] P.-L. Lions: Mathematical topics in fluid mechanics, Vol. I: incompressible models. Oxford Lecture Series in Mathematics and its applications, 3 (1996), Oxford University Press. [107] P.-L. Lions: Mathematical topics in fluid mechanics, Vol. II: compressible models. Oxford Lecture Series in Mathematics and its applications, 10 (1998), Oxford University Press. [108] P.-L. Lions: Sur les ´equations diff´erentielles ordinaires et les ´equations de transport. C. R. Acad. Sci. Paris S´er. I, 326 (1998), 833–838. [109] J. Lott & C. Villani: Weak curvature conditions and functional inequalities. J. Funct. Anal., in press. [110] S. Maniglia: Probabilistic representation and uniqueness results for measure-valued solutions of transport equations. J. Math. Pures Appl., 87 (2007), 601–626. [111] S. Maniglia: Fine properties and well-posedness of solutions of transport equations and of the 1st order associated ODE. PhD Thesis, Universit` a di Pisa, 2006. [112] J. N. Mather: Minimal measures. Comment. Math. Helv., 64 (1989), 375–394. [113] J. N. Mather: Action minimizing invariant measures for positive definite Lagrangian systems. Math. Z., 207 (1991), 169–207. [114] E. Y. Panov: On strong precompactness of bounded sets of measure-valued solutions of a first order quasilinear equation. Math. Sb., 186 (1995), 729–740. [115] G. Petrova & B. Popov: Linear transport equation with discontinuous coefficients. Comm. PDE, 24 (1999), 1849–1873. [116] F. Poupaud & M. Rascle: Measure solutions to the liner multi-dimensional transport equation with non-smooth coefficients. Comm. PDE, 22 (1997), 337–358. [117] A. Pratelli: Equivalence between some definitions for the optimal transport problem and for the transport density on manifolds. Ann. Mat. Pura Appl., 184 (2005), 215–238. [118] S. K. Smirnov: Decomposition of solenoidal vector charges into elementary solenoids and the structure of normal one-dimensional currents. St. Petersburg Math. J., 5 (1994), 841–867. [119] E. M. Stein: Singular integrals and differentiability properties of functions. Princeton University Press, 1970. [120] M. E. Taylor: Partial differential equations. Basic theory. Texts in Applied Mathematics, 23 (1996), Springer-Verlag, New York. [121] L. Tartar: Compensated compactness and applications to partial differential equations. Research Notes in Mathematics, Nonlinear Analysis and Mechanics, ed. R. J. Knops, vol. 4, Pitman Press, New York, 1979, 136–211. [122] R. Temam: Probl´emes math´ematiques en plasticit´e. Gauthier-Villars, Paris, 1983. [123] H. Triebel: Fractals and Spectra related to Fourier analysis and function spaces. Monographs in Mathematics 91, Birkh¨ auser Verlag, 1997.
142
BIBLIOGRAPHY
[124] H. Triebel & H. Winkelvoss: A Fourier analytical characterization of the Hausdorff dimension of a closed set and of related Lebesgue spaces. Studia Math., 121 (1996), 149–166. [125] J. I. E. Urbas: Regularity of generalized solutions of Monge–Amp`ere equations. Math. Z., 197 (1988), 365–393. [126] A. Vasseur: Strong traces for solutions of multi-dimensional scalar conservation laws. Arch. Rational Mech. Anal., 160 (2001), 181–193. [127] C. Villani: Topics in mass transportation. Graduate Studies in Mathematics, 58 (2004), American Mathematical Society. [128] C. Villani: Optimal transport: old and new. Lecture Notes of the 2005 Saint-Flour Summer school. [129] L. C. Young: Lectures on the calculus of variations and optimal control theory. Saunders, 1969. [130] W. P. Ziemer: Weakly differentiable functions. Sobolev spaces and functions of bounded variation. Graduate Texts in Mathematics, 120 (1989), Springer-Verlag.